SLI 3D Fingerprint Capture Specifications Page 1 ECE TECHNICAL REPORT CSP-12-007 Department of Electrical and Computer Engineering Center for Visualization and Virtual Environments University of Kentucky Test Procedures for Verifying Image and 3-Dimensional Quality Requirements For Personal Identity Verification (PIV) Single Fingerprint Structured Light Illumination (SLI) Device September (updated 6-28-2013) Laurence G. Hassebrook Document modeled after, and using, excerpts from: Norman B. Nill, “TEST PROCEDURES FOR VERIFYING IMAGE QUALITY REQUIREMENTS FOR PERSONAL IDENTITY VERIFICATION (PIV) SINGLE FINGERPRINT CAPTURE DEVICES,” MITRE Technical Report MTR060170R3, MITRE, Bedford Massachusetts, December 2006. DEFINITIONS Orange Peel Deformation: Visual description of what is technically known as map deformation. It is the stretching and/or compressing of a curved manifold when flattened to a 2-D surface. Skin Stretch Deformation: Also includes skin compression. The deformation of skin beyond the Orange Peel Deformation and related to the full structure of a finger including the surface friction ridges, the underlying flesh, muscle and bone and also the surface in contact with the finger as well as the pressure applied. Abstract We propose a certification modeled after the original 2-D certification. Our effort has been to Keep It Simple (KIS) and our philosophy is that if we can certify that the 3-D scanner yields quality 3-D data and the flattening plugin can approximate the orange peel and skin stretch distortions in a real finger, that we will have a fingerprint capture that is as good or better than the 2-D contact systems.
Transcript
SLI3DFingerprintCaptureSpecifications Page1
ECE TECHNICAL REPORT CSP-12-007 Department of Electrical and Computer Engineering Center for Visualization and Virtual Environments University of Kentucky Test Procedures for Verifying Image and 3-Dimensional Quality Requirements For Personal Identity Verification (PIV) Single Fingerprint Structured Light Illumination (SLI) Device September (updated 6-28-2013) Laurence G. Hassebrook Document modeled after, and using, excerpts from: Norman B. Nill, “TEST PROCEDURES FOR VERIFYING IMAGE QUALITY REQUIREMENTS FOR PERSONAL IDENTITY VERIFICATION (PIV) SINGLE FINGERPRINT CAPTURE DEVICES,” MITRE Technical Report MTR060170R3, MITRE, Bedford Massachusetts, December 2006.
DEFINITIONS
Orange Peel Deformation: Visual description of what is technically known as map deformation. It is the stretching and/or compressing of a curved manifold when flattened to a 2-D surface.
Skin Stretch Deformation: Also includes skin compression. The deformation of skin beyond the Orange Peel Deformation and related to the full structure of a finger including the surface friction ridges, the underlying flesh, muscle and bone and also the surface in contact with the finger as well as the pressure applied.
Abstract
We propose a certification modeled after the original 2-D certification. Our effort has been to Keep It Simple (KIS) and our philosophy is that if we can certify that the 3-D scanner yields quality 3-D data and the flattening plugin can approximate the orange peel and skin stretch distortions in a real finger, that we will have a fingerprint capture that is as good or better than the 2-D contact systems.
3.4 Test Procedures ................................................................................................................................ 13
3.4.1 Across‐Bar Geometric Accuracy and Resolution Scale .............................................................. 13
4.3 Test Surface ....................................................................................................................................... 18
4.3.1 Sine Wave Test Surface .............................................................................................................. 18
4.3.2 Bar Test Surface or 3‐D Ronchi Ruling ....................................................................................... 19
4.3.3 Step Edge Test Surface ............................................................................................................... 20
4.4 Test Procedures with Sine Wave Surface or 3‐D Ronchi Ruling ........................................................ 20
4.5 Test Procedures with Step Edge Surface .......................................................................................... 22
6.3 Test Procedures ................................................................................................................................ 31
6.3.1 Fingerprint Gray Range .............................................................................................................. 31
7. List of References ................................................................................................................................ 34
A. APPENDIX: FOURIER COMPONENT REPRESENTATION ....................................................................... 34
A.1 Fourier Component Analysis of Blurring with 2D and 3D data ......................................................... 35
A.2 Fourier Component Analysis of 3D Ronchi Ruling Surface ............................................................... 37
B. APPENDIX: FLATTENING DILEMMA; ORANGE PEEL AND SKIN STRETCH DEFORMATION .................. 41
C. APPENDIX: GRAY LEVEL ENCODING .................................................................................................... 44
C.2 Average and Modulation Intensity. .................................................................................................. 46
D. APPENDIX: RECTILINEAR COORDINATES ............................................................................................. 47
F. APPENDIX: APPENDIX F, PIV and SLI CROSS REFERENCING ................................................................ 49
SLI3DFingerprintCaptureSpecifications Page4
1. Introduction
1.1StructuredLightIllumination3‐DSensorTestOrganization The testing and specifications for non-contact Structured Light Illumination (SLI) are analogous to the contact based 2-D imaging systems. However, there are differences as shown in Fig. 1.1. The sensor itself, shown in Fig. 1.2, projects a series of image patterns which are processed temporally at each pixel location to yield a 3-D coordinate for each pixel in the sensor. For the testing procedure, the performance is based on the 3-D coordinate point cloud up to and including the Gray Level Uniformity as indicated in Fig. 1.1. For the most part, either the peak-to-peak pattern variation or the physical ridge depth is substituted in place of image intensity in the performance measurements. Starting at the “flattening” process, the 3-D data is flattened to 2-D via two non-linear operations; (1) orange peel flattening and (2) skin distortion. Once flattened it is expected that the non-uniform sample spacing is denser than 500 ppi. The final step is to interpolate the non-uniformly spaced samples to a uniformly spaced 500 ppi grid and stored as a gray level image. The goal of flattening is to be operable with other 3-D scanners, matching algorithms, legacy plane print data bases and human inspection.
Figure 1.1: 2‐D contact versus 3‐D non‐contact specification and testing organization.
SLI3DFingerprintCaptureSpecifications Page5
Figure 1.2: End‐to‐end SLI system including the flattening process.
1.2SLIIntensityEquivalents
Unlike traditional photography, SLI provides signal and depth information in image form. The reason is that the D1 SLI is implemented by projecting a series of sine wave patterns sequentially shifted in the “phase” direction. In this report, the phase dimension is vertical or along the Y dimension of the imagery. What we call the “orthogonal” dimension is horizontal or along the X direction. A set of projection patterns for a single cycle or unit frequency is shown in Fig. 1.3.
SLI3DFingerprintCaptureSpecifications Page6
Figure 1.3: Unit frequency pattern projection sequence.
The projected patterns are described in projector coordinates {xp, yp} as
NnyfBAyxI pcppppnoj 22cos,,Pr (1.1)
where in Fig. 1.3, fc=1 and N=8 and n=0, 1, 2,… , 7. The camera captures these patterns at an angle, reflected off the target surface and hence, distorted by the depth of the surface. The received images in camera space {xc, yc} are described by
cczccccccn yxNnyxByxAyxI ,2cos,,, (1.2)
Since the patterns occur in time, we can refer to the processed results from the patterns as an image signal in that each pixel has a time signal associated with it. There are 3 image types that we can calculate from Eq. (1.2). The first one is simply an average of all the patterns and looks like a photographic image of the target. If N is even and is 4 or larger, then the average intensity image signal is
1
0
,1
,N
nccncc yxI
NyxA (1.3)
The other two image signals are the intensity modulation, B(xc, yc), and the phase, z(xc, yc), image signals. The intensity modulation is a measure of peak-to-peak reflectance of the patterns and the phase is directly related to world coordinates via a perspective transformation. The measurable outputs from are the world coordinates {Xw(xc, yc), Yw(xc, yc), Zw(xc, yc)} where Zw(xc, yc) is used in place of intensity in some of the specifications. The intensity equivalent parameters of SLI are listed in Table 1-1.
SLI3DFingerprintCaptureSpecifications Page7
Table 1‐1: SLI intensity equivalents that may be used in place of intensity measures in performance specifications.
Intensity Equivalent Parameter
Description of Parameter
A(xc, yc) Intensity average of the patterns at each pixel location {xc, yc}. B(xc, yc) Intensity modulation indicates the peak-to-peak variation of the reflected patterns
at each pixel location {xc, yc}. Zw(xc, yc) Depth measure. After a flattening process is applied, the finger curvature is
removed but the ridge cross sections still have depth variation. These ridge depths are encoded with gray levels corresponding to an intensity equivalent.
1.3GeneralTestAttributes
Verification of compliance of a capture device to the 3D PIV spec requirements is primarily performed by the test method; i.e., verification through systematic exercising of the item with sufficient instrumentation to show compliance with the specified quantitative criteria. Some requirements as noted in the text, may be verified by the inspection method; i.e., verification of requirements by visual examination of the item or review of descriptive documentation [VERIFY].
All required testing is the responsibility of the vendor who is seeking certification and is
performed at a vendor-designated facility. The FBI neither performs the testing, nor witnesses the testing, nor receives the physical device being tested. However, the vendor must submit the digital test images to the FBI for independent analysis. These test images must be submitted in an uncompressed format1 such as raw, TIFF, BMP, or PGM, at 8 bits per pixel (8 bpp, 256 gray levels). The vendor may also supply a technical description of the capture device and a test report of its internal test results, but neither of these documents will substitute for submission of the test images.
It is expected that all testing will be performed on a single, representative unit of the product/model for which certification is being sought.
The device shall be tested to meet requirements in its normal-operating-mode, with the following possible exceptions:
1) If the device has a strong anti-spoofing feature, of a type whereby only live fingerprints will produce an image, then this feature needs to be switched-off or bypassed in the target test mode of operation.
2) If the device’s normal output is not a monochrome gray scale image, e.g., it is a binary image, minutia feature set, color image, etc., then the monochrome gray
1 Uncompressed images that were previously compressed via a lossy compression, such as WSQ or JPEG, are not acceptable.
SLI3DFingerprintCaptureSpecifications Page8
scale image (such as from an intermediate processing step) needs to be accessed and output in the test mode of operation.
3) If the device’s normal output is a 3-Dimensional point cloud followed by a flattening algorithm for backward compatibility then the 3-D data may be used to evaluate the sensor resolution, geometric accuracy, gray level uniformity and geometric accuracy prior to flattening. The flattening algorithm testing will be grouped with the test imaging, as shown in Fig. 1.1, and have its own set of performance tests and be part of the test image verification process in Section 6.
4) Other normal-operating-mode features of the device similar or analogous to (1), (2) and (3) may be disengaged.
Prior to submission of test data, the vendor should open a dialogue with the FBI and
describe the proposed test targets and test procedures, the goal being to obtain a consensus on suitable and acceptable targets and procedures. This is particularly important if the vendor has not previously obtained fingerprint capture device certification from the FBI utilizing the given imaging technology.
The vendor is encouraged to analyze its test data results before submission to the FBI; test data analysis software (freeware) is available through the FBI for this purpose.
If the test data analysis under the direction of the FBI indicates noncompliance with any of the PIV spec requirements, the vendor will be informed of the specific deficiencies and given an opportunity to correct the deficiencies and submit new test data.
SLI3DFingerprintCaptureSpecifications Page9
2. BasicRequirements The basic requirements for the PIV single finger capture device are given in Table 2-1.
Table 2‐1: Basic PIV Device Requirements
Parameter
Requirement
Capture Size
> 12.8 mm wide by > 16.5 mm high
True Optical or Native Resolution (Nyquist frequency)
> 500 ppi in sensor detector row and column directions.
Resolution Scale 490 ppi to 510 ppi in sensor detector row and column directions.
Image Type Capability to output monochrome image at 8 bits per pixel, 256 gray levels (prior to any compression).
mm = millimeters ppi = pixels per inch > greater than or equal to bpp = bits per pixel
2.1Background For 3-D SLI technology, all measurements are performed on a flat surface at depth Z0 where it is referenced to the extrema depth values as Zmin < Z0 < Zmax. The depth range Zmin to Zmax is established in Section 4.
The dimensions of the minimum capture size defined in Table 2-1 refers to a rectangle. If the capture area is some other shape, then the minimum size rectangle must fit within the boundaries of that shape, e.g., the minimum size square capture area would be 16.5 by 16.5 mm, or an elliptical capture area would need a minor axis width greater than 12.8 mm in order to fit a 12.8 mm wide rectangle within its boundary.
The true optical or native resolution of the device must be at least 500 ppi; if the
fingerprint is captured at a resolution level of, for example, 800 ppi, then that image is downsampled/rescaled to 500 ppi for output, using an appropriate rescaling technique. It is unacceptable to capture the fingerprint at any ppi level less than 500 ppi, and then upsample it to 500 ppi.
The PIV device’s aimpoint resolution scale should be 500 ppi, in which case the actual
resolution scale might vary between 490 and 510 ppi, due, for example, to small
SLI3DFingerprintCaptureSpecifications Page10
directional or field angle dependent imaging perturbations. The aimpoint resolution scale could be greater than 500 ppi, up to 510 ppi, but should never be less than 500 ppi.
All test target and test fingerprint scans are output as 8 bpp, 256 gray level monochrome
images, irregardless of the device’s operational fingerprint image type output. Test fingerprint and operational fingerprint image polarity is expected to be: dark gray ridges with light gray valleys.
2.2RequirementsCompliance
Verification that the device’s capture size contains a rectangle of at least 12.8 mm wide by 16.5 mm high is performed by noting the fingerprint image width and height in pixels and dividing by the independently measured horizontal and vertical direction resolution scales, respectively, in pixels per inch.
Verification that the device’s true optical or native resolution is at least 500 ppi is
performed by the SFR assessments in vertical and horizontal directions2 (see section 4), and any available relevant documentation on the device, such as number of independent sensor pixels spanning the vertical and horizontal directions of the capture area. In the case of 3-D SLI technology, this verification is performed on a flat surface at depth Z0.
Verification that the device’s resolution scale is between 490 ppi and 510 ppi, in both
vertical and horizontal directions, is performed by the geometric accuracy measurements (see section 3) or, if none, from measurements on the image of the SFR target, in the vertical and horizontal directions, i.e., image pixel distance divided by corresponding target distance in inches.
Verification that the device’s image type output is 8 bits per pixel is performed by
inspection of the test images to confirm that they are monochrome, together with results of the fingerprint gray range assessment (see section 6).
2.3Discussionof2Dand3DSamplingUniformity We recommend having the same ppi specification for both contact 2D and noncontact 3D SLI technology. That is, this specification in the 3D SLI device is to choose one focal plane within the depth range of the device and achieve 500 ppi on that plane. The justification is that the we believe the spirit of this specification is not to be as rigorous as the other specifications in later sections but to simply categorize the technology such whether it is “500 ppi” or “1000 ppi” type of technology. Furthermore, this specification brings up some interesting discussion about how uniform the sampling in 2D and 3D really is. As it turns out, when a finger is pressed against a flat surface, 2 All references to vertical and horizontal directions in this document assume that the sensor detector columns and rows, respectively, align with these image directions. If such is not the case, then ‘sensor detector columns and rows’ directions replaces all mention of ‘vertical and horizontal’ directions.
SLI3DFingerprintCaptureSpecifications Page11
the ridges deform from their natural 3D position. This means that even in the 2D contact case, the sampling is not really uniform because some ridges are made wider or narrower depending where they are on the flat print. In the case of 3D SLI capture, the deformation can be accurately characterized and occurs in the original camera images because of the finger surface curvature so the distortion increases toward the outer regions of fingerprint. The flattening dilemma due to orange peel and skin stretch deformations is discussed in detail in Appendix B. The underlying cause of the flattening dilemma also affects sampling uniformity in both flattened 3-D prints as well as 2-D contact prints. Appendix B presents our argument for keeping the specification for 500 ppi in Section 2 to be as simple for 3-D SLI as it is for 2-D contact fingerprint devices.
3. GeometricAccuracy Geometric Accuracy of an SLI 3-D sensor with flattening turned off is measured exactly the same as a 2-D contact with the exception that the Ronchi ruling must be matte surface of alternating black and white reflectance and it is precisely positioned at 3 different height levels to allow measurement of “flatness” along the Z-dimension and an accurate absolute Z height value. The Z-dimension flatness is measured by fitting a plane to the white Ronchi ruling bars at the 3 different height levels. The accuracy of the flatness will most likely be half that of the straightness in the X-Y plane. This is primarily due to the angle between the projector and camera and is inherent in SLI devices. This lower performance should not be an issue considering that the area of the fingers being processed will not have large gradients, much beyond what 2-D contact prints do. The final error from the depth error introduced into the X-Y directions after flattening can be shown to be negligible.
3.1Requirements Requirement #1 (across-bar) A multiple, parallel bar target with a one cy/mm frequency is captured in vertical bar and horizontal bar orientations. The absolute value of the difference between the actual distance across parallel target bars, and the corresponding distance measured in the image, shall not exceed the following values, for at least 99% of the tested cases in each of the two orthogonal directions.
D ≤ 0.0013, for 0.00 < X ≤ 0.07 D ≤ 0.018X, for 0.07 ≤ X ≤ 1.50
where D=|Y-X|, X= actual target distance, Y=measured image distance and D, X, Y are in inches. Requirement #2 (along-bar):
SLI3DFingerprintCaptureSpecifications Page12
A multiple, parallel bar target with a one cy/mm frequency is captured in vertical bar and horizontal bar orientations. The maximum difference between the horizontal direction locations (for vertical bar) or vertical direction locations (for horizontal bar), of any two points separated by up to 1.5 inches along a single bar’s length, shall be less than 0.027 inches for at least 99% of the tested cases in the given direction.
3.2Background Across-bar geometric accuracy is measured across the imaged Ronchi target bars, which must cover the total image capture area. The requirement corresponds to a positional accuracy of ± 1.8% for distances between 0.07 and 1.5 inches, and a constant ± 0.0013 inches (2/3 pixel at 500 ppi) for distances less than or equal to 0.07 inches. These across-bar measurements are also used to verify compliance with the device’s resolution scale tolerance requirement given in Table 2-1. Along-bar geometric accuracy is measured along the length of individual Ronchi bars in the image. For a given horizontal bar, for example, the maximum difference between bar center locations (in vertical direction), determined from bar locations measured at multiple points along the bar’s length, is compared to the maximum allowable difference requirement (analogously for vertical bar). This requirement is to ensure that pincushion, barrel, or other types of distortion are within acceptable bounds.
3.3Target The multiple, parallel bar target refers to a Ronchi target, which consists of a series of parallel black bars with white/clear spaces between, at a spatial frequency of 1.0 cy/mm; i.e., each black bar and each white/clear space between two adjacent bars has a width of 0.50 mm, as illustrated in Fig. 3.1. Ronchi targets are commercially available on a number of substrates, e.g., transparent film, reflective mylar, and as a chrome-on-glass target. However, it is the PIV vendor’s responsibility to identify a 1.0 cy/mm Ronchi target that will produce a suitable, measurable image in the vendor’s PIV device; this may entail experimentation and design/fabrication of a custom Ronchi target.
The Ronchi target must cover the entire capture area with vertical bars in one image and horizontal bars in another image, each aligned to within 0.5 degrees of the relevant axis (vertical or horizontal). Acquiring a single image containing both vertical and horizontal bars, such as in a cross-hatched pattern or segmented vertical and horizontal bars, is unacceptable. If the PIV device operates under optical imaging principles then, in order to achieve more uniform contact, it is acceptable to use an index-of-refraction matching liquid between the target and the device’s fingerprint platen.
SLI3DFingerprintCaptureSpecifications Page13
Figure 3.1: 1.0 cy/mm Ronchi Target (Covers the Example 15.5 x19 mm Device Capture Area)
3.4TestProcedures The following outlines the procedures for measuring the PIV device’s geometric accuracy and resolution scale; these measurements are performed by the geo software application, supported by creategeofile and viewgeo applications, which are on the Test Tools CD (see Appendix A of reference [1] ). A more detailed description of the computations in geo is given in Appendix B of reference [1]. 3.4.1Across‐BarGeometricAccuracyandResolutionScale The goal is to acquire measurements that comprehensively cover a Ronchi target image, in continuous, 0.25 inch wide measurement strips running the height (vertical direction measurements) or width (horizontal direction measurements) of the device capture area. If imaging artifacts caused by target blemishes, dust, target-device contact non-uniformities, etc. prevent successful measurement in some isolated areas, then measurement locations would need to be shifted. For each measurement strip, measurements are taken across all independent 1-bar cycles (1-bar distance) and 6-bar cycles (6-bar distance). All distances for geometric accuracy are measured from bar center to bar center in pixel units, as illustrated in Figure 3-2. A bar center is located by first detecting that bar's left and right edges along a 0.25 inch edge height (for vertical bar), and then bisecting the bar’s two edges, taking skew angle into account. The device’s output resolution scale is first measured over 6-bar cycles, which is used to verify compliance with the pixels per inch resolution scale requirement and establish a pixels per inch value that can be used to convert subsequent geometric accuracy measurements to inches. Next, measurements are made to test the geometric accuracy over a short distance of 1-bar cycle, and then over the longer distance of 6-bar cycles. The pixel measurements are then converted to inches using the previously computed average resolution for the given measurement strip.
SLI3DFingerprintCaptureSpecifications Page14
Figure 3.2: Six‐Bar and One‐Bar Distance Measurements.
3.4.2Along‐BarGeometricAccuracy This test for distortion utilizes the same Ronchi target images used for across-bar geometric accuracy assessment. All distances for along-bar geometric accuracy are first measured from local bar center location to local bar center location in pixel units. The difference in fractional rows between two bar centers is converted to inches by dividing by the average ppi of the two corresponding measurement strips. Fig. 3.3 illustrates a single measurement for a single horizontal bar in a device that has a 1.5 inch capture width. In this example, the maximum vertical deviation “H” occurs for the bar center in center field compared to the bar center at the edge of the field, but note that the maximum vertical deviation could occur at other bar center location pairs. The vertical deviation shall not exceed 0.027 inches per 1.5 inch length.
Figure 3.3: Along‐Bar Geometric Accuracy Assessment (Vertical Height "H" is Test Sample Value)
3.5RequirementsCompliance The across-bar geometric accuracy requirement is complied with if at least 99 percent of the tested cases are within the minimum and maximum distance limits given in Table 3-1, in the vertical and horizontal directions. Table 3‐1: Geometric Accuracy Requirements
Measurement Correct Value Directional Requirement Met if: 1-bar distance 0.03937 > 99% in range: 0.03807 to 0.04067 6-bar distance 0.23622 > 99% in range: 0.23197 to 0.24047 The along-bar geometric accuracy requirement is complied with if at least 99 percent of the test measurement values (“H” in Figure 3-3) are less than 0.027 inches, in the vertical and horizontal directions.
SLI3DFingerprintCaptureSpecifications Page15
The resolution scale requirement is complied with if the average resolution scale is between 490 ppi and 510 ppi in the vertical and horizontal directions (see section 2).
3.6No‐TestOption See reference [1]. This specification is not applicable to SLI. The across-bar and along-bar geometric accuracy requirements defined in section 3.1 may be verified by the Inspection Method3 instead of the Test Method, if the fingerprint capture device has all 5 of the following characteristics, and adequate documentation for each of these characteristics is submitted to the FBI.
3.7DepthResolutionandNoiseLevel From the placement of the flat targets at the 3 depth levels we can establish the depth accuracy. We include a specific measure of depth noise and an analogous measure of depth linearity with any one of the following 3 measurements.
1. Use a stair case height target that has steps between Zmin and Zmax at 40 micron increments. The step heights could be used to fit a line and measure depth accuracy.
2. A slanted plane can be used to measure depth linearity. A plane would be fit to the plane to estimate flatness and the Standard Deviation from the plane.
3. A precision translation stage can be used to move a flat surface to different heights to evaluate accuracy of depth. Warning: this technique may not work with some SLI algorithms that measure relative depth rather than absolute.
We base the threshold on mapping the depth to gray level. At this time there is no specification threshold on depth. In appendix F the specification is for gray level and is in section 2.1 of appendix F. In SLI depth can be treated as a spatial dimension or mapped to a gray level. Consider 7.65/255 as a threshold scaling coefficient based on gray level mapping. Then for a 5 mm depth span, the allowed variation would be 0.15 mm =5mm x 7.65/255. 3.8StructuredLightIlluminationDepthBandingTest A common unwanted artifact with SLI pattern projection techniques is known as “banding. There are several sources of banding including gamma distortion, motion banding and pixel banding. Gamma distortion is caused by sine wave patterns being non-linearly distorted based on a parameter known as gamma. Motion during the scanning of the projection patterns shifts the patterns and pixel banding occurs when the camera and optical resolution is high compared to the projector resolution such that the camera actually captures images of the projector pixels. All of these distort what is known as the phase image in a periodic manner in the direction of the lateral depth distortion or “phase dimension”. The banding parameter that is important to measure is the peak-to-peak (p-p) variation in depth. We recommend that the p-p variation be below 15 microns.
SLI3DFingerprintCaptureSpecifications Page16
4. SpatialandDepthSpatialFrequencyResponse(DSFR) For a SLI sensor, there are at least three different outputs, as described in subsection 1.2 that can be used for frequency response analysis. The first two are treated as intensity and follow the same testing procedures as defined in references PIV [1] and/or appendix F [5]. For the 3rd depth parameter, the analogy to detected intensity is the detected 3-D depth. However, the lateral coordinates of a 2-D image are in pixel units (sometimes referred to as “uv” space or camera space) the lateral units of an SLI scan are in world coordinates {Xw, Yw} with the depth also in {Zw} and units of millimeters. Therefore, the range of the depth modulation z in a 3-D Ronchi ruling as well as a 3-D sinusoidal surface must be set by the specification to be practical. Given a practical value of depth modulation z, this analogy allows the value of depth to be substituted in place of captured intensity in determining the “Depth” Spatial Frequency Response (DSFR) parameters of “D” CTF (DCTF) and “D” MTF (DMTF).
4.1Requirements The spatial depth-frequency response shall normally be measured by either using a bi-level bar target (i.e., 3-D Ronchi ruling), which results in the device’s Depth Contrast Transfer Function (DCTF) or by using a sinusoidal target surface, which results in the device’s Depth Modulation Transfer Function (DMTF). If the device cannot use a bar target or sine wave target, i.e., a useable/measurable surface cannot be produced with one of these target surfaces, then an step edge target can be used to measure the DMTF3. All target surfaces should have uniform albedo, matte reflective properties and be opaque with no internal scattering. The DCTF or DMTF shall meet or exceed the minimum modulation values defined in Eq. (4-1) and Table 4-1 (for DCTF) or Eq. (4-2) and Table 4-2 (for DMTF), over the frequency range of 1.0 to 10.0 cy/mm, in both the detector row and detector column directions, and over any region of total capture volume. Table 4-3 gives the minimum DCTF and DMTF modulation values at nominal test frequencies. None of the DCTF and DMTF modulation values in 1.0 to 10.0 cy/mm range shall exceed 1.12, and the target image shall not exhibit any significant amount of aliasing in that range. The performance threshold for the DCTF is given as:
4
0n
nnDCTF fafT (4-1)
where the coefficients are Table 4‐1: Coefficients for DCTF performance threshold.
Coeff. Name Coefficient Value a0 1.00838 a1 -8.05399 x 10-2 a2 -8.94631 x 10-3
3 If it is conclusively shown that neither a sine wave target surface, nor bar target surface, nor edge target surface can be used in a particular device, other methods for DSFR measurement may be considered.
SLI3DFingerprintCaptureSpecifications Page17
a3 +1.43781 x 10-3 a4 -5.71711 x 10-5 and for DMTF the performance threshold function is
3
0n
nnDMTF fbfT (4-2)
where the coefficients are Table 4‐2: Coefficients for the DMTF performance threshold.
Coeff. Name Coefficient Value b0 1.02829 b1 -1.67473 x 10-1 b2 +1.06255 x 10-2 b3 -2.80874 x 10-4 The range of f is from 1 to 10 cy/mm and the resulting performance thresholds are shown in Table 4-3. 4‐3: Minimum DCTF and DMTF Requirements at Nominal Test Frequencies.
f(cy/mm) Minimum DCTF Minimum DMTF
1.0 0.920 0.871
2.0 0.822 0.734
3.0 0.720 0.614
4.0 0.620 0.510
5.0 0.526 0.421
6.0 0.440 0.345
7.0 0.362 0.280
8.0 0.293 0.225
9.0 0.232 0.177
10.0 0.174 0.135
4.2BackgroundThe 1.12 upper limit for modulation is to discourage image processing that produces excessive edge sharpening, which can add false detail to an image and/or excessive noise. This specification should work in the same way with 3-D information. The target should be fabricated from opaque and matte reflective material/s. While human fingers are curved surfaces, for testing purposes, we want to isolate the spatial frequency performance as it is throughout the scan volume of the 3-D fingerprint scanner. Thus, we have chosen to use “flat” test surfaces with local depth variations for measuring the DCTF and DMTF rather than curved surfaces which do not yield a uniform sampling of the scan volume as does multiple measurements of the flat test surfaces at different depths.
SLI3DFingerprintCaptureSpecifications Page18
See additional 2-D contact specification discussion in reference [1].
4.3TestSurface It is not required that the DCTF or DMTF be obtained at the exact frequencies listed in Table 4-3; however, the DCTF or DMTF does need to cover the listed frequency range, and contain frequencies close to each of the listed frequencies. For any bar target surface used, or a custom sine wave target, the target frequency patterns must be within 0.49 cy/mm of each of the frequency increments: 1, 2, 3, 4, 5, 6, 7, 8, 9 cy/mm, and within 0.25 cy/mm of 10 cy/mm. Also, a minimum number of sine wave cycles or bars are needed at each frequency, in order to: (1) ensure capturing the optimum phase between the device’s sensor pixels/points and the target sine waves or bars, (2) have enough samples available for accurate measurement of aliasing, and (3) obtain an accurate measure of modulation. Table 4-4 specifies the minimum number of sine wave cycles or bars as a function of frequency. Table 4‐4: Minimum Number of Target Bars or Sine Wave Cycles (500 ppi Device)
Target Frequency (f) Minimum Number of Bars or Sine Wave Cycles 1 cy/mm 4 > 1 to 4 cy/mm 5 > 4 to 10 cy/mm 10 Notes Bar Target: bar length must be at least 10 times bar width; width of space between parallel bars equals bar width. Sine Target: length of sine wave in direction perpendicular to sinusoidal variation must be at least 5/f.
4.3.1SineWaveTestSurface At this time we have not found a source for accurate 3-D sine wave surfaces. So our discussion will continue in the next Subsection with the 3-D Ronchi ruling.
SLI3DFingerprintCaptureSpecifications Page19
Figure 4.1: (a) Physical section of a typcial sine wave target in 2‐D. (b) 3‐D sine wave mapped to flat test surface. Note that the periods are the same for both test grids but the depth variation is analogous to the reflectance variation in the 3‐D and 2‐D grids.
4.3.2BarTestSurfaceor3‐DRonchiRuling
SLI3DFingerprintCaptureSpecifications Page20
At microscopic scale, 3-D sine wave surfaces are difficult to make so an alternative test surface is a 3-D Ronchi ruling. The 3-D Ronchi ruling can be made by using a micro-milling machine or other micro-machining or photographic techniques. For example, the material may be metallic and roughened by anodization or etching to have a matte reflectivity (this will be updated as we get the surfaces made or find one already available). A comparison of a 2-D Ronchi ruling with a 3-D Ronchi ruling is shown in Fig. 4-2 (a) and Fig. 4-2 (b), respectively. A different test surface for each frequency is used in the testing and moved along the Z direction with precision translation stages or positioning blocks to place the test surface at different depths in the scan volume as shown in the SLI configuration of Fig. 4.3. The z is the same for all grid frequencies for practical reasons but can be normalized for each grid if the grids used are different in z peak-to-peak (p-p) depth variation. We recommend z = 50 microns so that at f=10 cy/mm the width of the lines would be equal to the height.
Figure 4.2: (a) 2‐D Ronchi ruling and (b) 3‐D Ronchi Ruling. Both have same period. See Appendix A for a compact specification for a composite test target.
4.3.3StepEdgeTestSurface Not discussed at this time but may be the easiest test surface to obtain. The Step edge surface would need to be positioned in more locations than the 3-D Ronchi ruling. For now, we will focus our attention on the 3-D Ronchi ruling.
4.4TestProcedureswithSineWaveSurfaceor3‐DRonchiRuling In Fig. 4-3, a typical SLI configuration is shown with a projector, which projects a series of encoded patterns, and a camera which captures the resulting reflected images containing the patterns laterally displaced by the surface depth contours. See reference [2] for GIF movies of pattern projections. The test procedure involves accurately positioning the test surface at different depths within the user defined scan volume of the system. This positioning is most likely accomplished by a precision translation stage which can be manual or motorized. The positional extrema of the scan volume must satisfy the lateral specifications of the PIV specifications [1] and extend in depth between Zmin and Zmax. Assume for discussion that z = 40 microns for all grid frequencies.
SLI3DFingerprintCaptureSpecifications Page21
Figure 4.3: SLI sensor configuration demonstrating the DSFR measurement at two different depths within the scan volume.
The pattern displacement due to depth is based on the angle between the projector and camera. Also, the reflected image is a function of both the camera and projector angles to the surface. Thus, there is an inherent asymmetry in SLI. For the measurement to take into account this asymmetry, the grid should be rotated in plane by 90 degrees and the measurements repeated. The resulting data set would have the test surface bars be both in the direction of the illumination as well as orthogonal to it. The following outlines the procedures for computing the PIV device’s DMTF from a sine wave target, or DCTF from a bar target. A Fourier component approach is used and a detailed description of the computations is in Appendix A with support documentation in reference [3].
1) A target data file is prepared containing information on the layout and dimensions of the target components, sine wave surface or 3-D Ronchi ruling surface modulation values, and step edge information.
2) The target surface is positioned on the device’s platen and aligned to within 0.5 degrees
of the horizontal and vertical, i.e., with respect to the device’s sensor row and column directions.
SLI3DFingerprintCaptureSpecifications Page22
3) The target surface is captured with the PIV device, after adjusting the device parameters (such as illumination, gain settings, etc.) to ensure that the projected sine wave pattern4 peaks and valleys are not at saturation.
4) A harmonic analysis application is run, with inputs consisting of 3-D coordinates along
pixel quality image, and various user-selectable options displayed at runtime.
5) The harmonic analysis application first computes the directional ppi scale and target alignment angle; this information is used together with the input data file to locate the row and column of each sine wave (or bar) frequency pattern and (for sine wave) the density patch locations, within the point cloud. The single, representative sine wave (or bar cycle) modulation in each captured 3-D frequency pattern is then determined from the sample modulation values collected from within that captured 3-D pattern. It is assumed that the X,Y,Z data are in units of mm. For the first order component of the pattern the DMTF is
00 fz
fz
M
minimummaximumfDMTF
pp
ppzz
(4.3)
where M0 = zp-p(f0) is the peak-to-peak modulation of the lowest frequency (i.e., 1 cy/mm) surface pattern. The direct measurement of the peak-to-peak height can be made but, for the 3-D Ronchi ruling, would yield the DCTF( f ) instead of the DMTF( f ) function. If each grid has a different z, then each DMTF( f ) function would be scaled by its true peak-to-peak divided by the lowest frequency peak-to-peak value.
6) Aliasing is also measured in harmonic analysis application, because it is a potential source
of unwanted depth artifacts, e.g., pronounced aliasing may produce false detail in the surface depth, such as a pseudo-ridge pattern. Aliasing can also indicate that the imaging device is operating in an unacceptable mode. Based on the concepts Fourier components in Appendix A and on reference [1] Appendix C, Aliasing is measured by first computing a sequence of one-dimensional discrete Fourier transforms (DFT) of the row-averaged depth levels in each frequency pattern, then inspection of the relative strengths of the DFT side lobes (harmonics) indicates the degree of aliasing-from-decimation, and the location of the DFT main lobe indicates whether or not aliasing-from-upscaling is present, which would be unacceptable. More discussion of aliasing detection is given in reference [1], Appendix C.
4.5TestProcedureswithStepEdgeSurface Not discussed at this time but may be the easiest to construct.
4.6RequirementsCompliance
4 Note that this “sine wave pattern” corresponds to the sequence of patterns that are projected onto the surface as shown in reference [2].
SLI3DFingerprintCaptureSpecifications Page23
The SFR requirement for a 500 ppi PIV device is complied with if, in both the detector (sensor) row and column directions, and over any region of the total capture area: (1) the measured DCTF is no less than that defined by equation 4-1, or the measured DMTF is no less than that defined by equation 4-2 (whichever applies), at each measured frequency in the range of 1.0 to 10.0 cy/mm and, (2) the measured DCTF or DMTF (whichever applies) is no greater than 1.12 at any frequency in the range of 1.0 to 10.0 cy/mm with z approximately 50 microns and, (3) any detected aliasing-due-to-decimation is within acceptable limits, where the acceptable amount decreases with decreasing frequency, reaching zero acceptable amount for frequencies less than 7cy/mm and,
(4) no aliasing-due-to-upscaling is detected at any frequency in the range of 0.0 to 10.0 cy/mm.
5. GrayLevelUniformity For the FS3D D1 device, the gray level coding is dependent on which of the image coding is used. The average point value across patterns and the amplitude modulation are treated the same as the contact scanner intensity values are.
If the depth value is used in place of intensity then the mapping to intensity is directly proportional to maximum and minimum depth in small regions that are large enough to contain multiple ridges. Thus, the ridges will be encoded to 255 and the valleys would have a minimum value of 0. Since the algorithm scales the range of Z depth values in a local area between 0 and 255, to measure the precision of the gray level uniformity it is easiest to work with the 3-D data directly. Two depth values of a single flat matte target plate is captured. Because it is possible to have a slightly slanted plate, a single flat plane can be calculated from the data and then based on its normal vector, the two scans are mathematically tilted to be level in Z depth. The two surfaces are mapped so their average Z depth is equal to the “light gray target” and “dark gray target” values given in the 2-D contact specifications.
The remainder of the testing is then performed identically to the 2-D contact specifications for “Adjacent Row,Column Uniformity,” “Pixel-to-Pixel Uniformity,” “Small Area Uniformity” and “Noise” as given in the 2-D contact specifications of references PIV [1] or Appendix F [5].
At least 99% of the average gray levels between every two adjacent quarter-inch long rows and 99% between every two adjacent quarter-inch long columns, within the capture area, shall not differ by more than 1.5 gray levels when scanning a uniform dark gray target, and shall not differ by more than 3.0 gray levels when scanning a uniform light gray target. #2 - Pixel-to-Pixel Uniformity For at least 99.0% of all pixels within every independent 0.25 by 0.25 inch area located within the capture area, no individual pixel's gray level shall vary from the average by more than 8.0 gray levels when scanning a uniform dark gray target, and no individual pixel's gray level shall vary from the average by more than 22.0 gray levels when scanning a uniform light gray target. #3 - Small Area Uniformity For every two independent 0.25 by 0.25 inch areas located within the capture area, the average gray levels of the two areas shall not differ by more than 3.0 gray levels when scanning a uniform dark gray target, and shall not differ by more than 12.0 gray levels when scanning a uniform light gray target. #4 - Noise The noise level, measured as the standard deviation of gray levels, shall be less than 3.5 in every independent 0.25 by 0.25 inch area located within the capture area, when scanning a uniform dark gray target and a uniform light gray target [1].
5.2Target A uniform light gray target and a uniform dark gray target are produced on any substrate that is compatible with imaging in the given PIV device. For example, reflective and transmissive uniform gray targets are commercially available; see Appendix A for sources. Alternatively, pseudo-targets may be substituted for physical targets. The pseudo-target concept images the blank capture area with the sensor exposure turned up or down (but not off!), by varying exposure time or illumination/excitation. Each target or pseudo-target must cover the entire capture area of the PIV device.
5.3TestProceduresandRequirementsCompliance The following gives the procedures for measuring the PIV device’s uniformity, with respect to requirements 1, 2, 3, and 4. These measurements are performed by MITRE‟s snr software application, which is on the Test Tools CD (see Appendix A); snr requires that the light gray image and dark gray image have the same size (width and height). The PIV device is adjusted so that the average image gray level from the light gray target is at least 4.0 gray levels below the maximum image gray level attainable with the device, and the average image gray level from the dark gray target is at least 4.0 gray levels above the minimum image gray level attainable with the device. For a full 8 bpp range image (0 to 255 gray levels), this implies that the light gray image is less than or equal to 251.0 and the dark gray image is greater than or equal to 4.0. However, if the device is setup such that some gray levels cannot
SLI3DFingerprintCaptureSpecifications Page25
occur, e.g., due to saturation or clipping, then the maximum or minimum value must be adjusted accordingly. For example, if gray levels 253, 254, 255 can never occur, then the maximum gray level would be 252 and the light gray image would need to be less than or equal to 248.0. Ideally, the average light gray image would be near the gray level of imaged fingerprint valleys and the average dark gray image would be near the gray level of imaged fingerprint ridges, that are typically acquired with the device (but still holding to the 4 gray level delta from maximum and minimum attainable gray levels). In all cases the polarity of the light and dark gray images are expected to be the same as for fingerprint scans, i.e., light uniform gray image corresponds to fingerprint valleys and dark uniform gray image corresponds to fingerprint ridges. Since these uniformity tests involve pixel-to-pixel and small-area difference measurements, foreign artifacts such as dust, target scratches, bubbles, or pinholes can adversely affect the results. Care in testing needs to be taken, e.g., use quality, uniform targets and clean the device’s platen. If a small quantity of measurement samples still contain residual foreign artifacts, for which it can be ascertained (such as by visual inspection of the image) that they are not part of the device, then these samples may be discounted from the final results. Due to the fact that single finger device capture areas are relatively small and are generally not a multiple of 0.25 inches in width or height, adhering to strict independence of all 0.25 inch measurement windows, row or column segments, could result in a substantial percentage of the capture area not being measured. To overcome this problem, it may be necessary to slightly overlap some of the adjacent 0.25 inch windows, row or column segments that are used in the various measurements. For example, if the width of the capture area is 375 pixels, then exactly 3 independent quarter-inch windows (each 125 pixels wide at 500 ppi) will fit across the width. However, if the capture area width is 350 pixels, then the leftmost and center windows can still be independent of each other, but the center and rightmost windows would have to overlap by 25 pixels in order to cover the capture width; with no overlap, only 2 windows across could be measured in this case and 29% of the width (100 columns) would not be measured. [The snr application automatically computes the overlap when necessary.]
5.3.1AdjacentRow,ColumnUniformity(Requirement#1) The average pixel gray levels of all 0.25 inch long horizontal row segments and all 0.25 inch long vertical column segments are computed over the entire capture area, for the light gray and dark gray images. The magnitude of the difference between the average gray levels of every two adjacent row segments and every two adjacent column segments are computed for each image. Figure 5.1 illustrates the measurements for row segments. Compliance with the requirement is achieved when at least 99.0% of the differences between adjacent row segment averages, and at least 99.0% of the differences between adjacent column segment averages, are less than or equal to 1.5 gray levels in the dark gray image, and are less than or equal to 3.0 gray levels in the light gray image.
5.3.2Pixel‐to‐PixelUniformity(Requirement#2) The locations of the minimum number of quarter-inch windows that cover the capture area are identified for the light gray and dark gray images. The average gray level for each quarter-inch window, rounded to the nearest whole number (nearest integer value), is computed for each image. The absolute value of the difference between the average of a given quarter-inch window and each of the individual pixel gray levels within that window is computed. Figure 5.2 illustrates the procedure (a 500 ppi device has 125 x 125 = 15,625 test values in each quarter-inch window). Compliance with the requirement is achieved when no more than 1.0% of the pixels in each dark gray image window are more than 8 gray levels away from the window average, and no more than 1.0% of the pixels in each light gray image window are more than 22 gray levels away from the window average.
5.3.3SmallAreaUniformity(Requirement#3) The locations of the minimum number of quarter-inch windows that cover the capture area are identified, for the light gray and dark gray images. The average gray level for each quarter-inch window is computed for each image (without rounding to whole number) as illustrated in Figure 5.3.
The absolute value of the difference between the gray level averages for every possible pairing of quarter-inch windows is computed. Compliance with the requirement is achieved when the largest difference for the dark gray image is less than or equal to 3.0 gray levels, and the largest difference for the light gray images is less than or equal to 12.0 gray levels.
SLI3DFingerprintCaptureSpecifications Page28
Figure 5.3: Area‐to‐Area Uniformity Measurements. Image from ref. [1].
5.3.4Noise(Requirement#4) The locations of the minimum number of quarter-inch windows that cover the capture area are identified, for the light gray and dark gray images. The standard deviation of pixel gray levels () is computed within each quarter-inch window in the light gray and dark gray images, where,
1
2
n
GGi (5.1)
where G = average gray level in given window, iG = gray level of ith pixel in given window and n = number of pixels in given window. Compliance with the requirement is achieved when the standard deviation of pixel gray levels
() is less than 3.5 in every 0.25 by 0.25 inch window located within the capture area, when scanning a uniform dark gray target and a uniform light gray target.
SLI3DFingerprintCaptureSpecifications Page29
5.4MeasurementofDeviceInput‐OutputRelation The input - output relation needs to be measured, either to ‘prove’ that the relation is linear, or to establish the actual nonlinear relation which would then need to be explicitly taken into account when computing the MTF or CTF. Since a COTS sine wave target incorporates its own step tablet; the mtf application can compute the input-output relation directly from this step tablet and take it into account in the sine wave MTF computations (whether it is nonlinear or linear). The input-output relation must be independently measured when using a bar target or edge target which does not contain a step tablet component. Measurement can be performed, for example, by inserting a series of neutral density filters into the imaging path, or by similarly varying the sensor exposure in discrete, known steps while imaging the blank platen, or by imaging a physical, pre-calibrated step tablet. Whatever method is used, it must produce enough steps to adequately define the input - output relation over the range of gray levels exhibited by the imaged sine wave, bar, or edge target. For example, Figure 5.4 illustrates a nonlinear input - output relation used with an edge target, where the edge image maximum gray and minimum gray levels are within the step tablet measurement range.
Figure 5.4: Example Nonlinear Input‐Output Relation.
6. FingerprintImageQuality For flattened 3-D fingerprint image quality, we use all the same specifications as used for the 2-D contact prints, except we add a geometric comparison between the legacy flat print minutia
SLI3DFingerprintCaptureSpecifications Page30
and the flattened 3-D prints. The company would provide all the required prints as well as minutia locations with correspondence between the flattened 3-D and 2-D contact prints. UK will provide a simple open source viewer so the examiners can see the superimposed minutia points and evaluate the alignment differences visually as well as a RMS difference value calculated by the UK viewer. If there are alignment problems, or if the company biased the minutia locations in their favor, the examiner could quickly tell that there is a problem and report objectively what the problem is. An example screen from this software is shown below. Note the RMS error is in microns and the data used is from 2008, binarized and would pass the specification with 300+ micron RMS error. We will need to determine a threshold for the RMS error performance.
Figure 6.1: (left) Flattend 3‐D and (2nd from left) 2‐D contact print with minutia defined. (right 2 images) Show the aligned and superimposed minutiae and also the far right image allows the inspector to click and drag a partition across to surface for comparison.
Figure 6.2: The Dashboard has information the inspector can see or the results can be opened up in the dialog box above.
SLI3DFingerprintCaptureSpecifications Page31
6.1Requirements Requirement #1 - Fingerprint Gray Range: At least 80.0 % of the captured individual fingerprint images shall have a gray-scale dynamic range of at least 150 gray levels. Requirement #2 - Fingerprint Abnormalities: The images must be true representations of the input fingerprints. Artifacts, anomalies, false detail, or cosmetic image restoration effects detected on the fingerprint images, which are due to the device or image processing, shall not significantly adversely impact supporting the intended applications. Requirement #3 - Fingerprint Sharpness and Detail Rendition: The sharpness and detail rendition of the fingerprint images, due to the device or image processing, shall be high enough to support the intended applications.
6.2Target The vendor submits a set of 20 finger images captured with the SLI device, nominally acquired from 10 different subjects, with 2 fingers per subject (preferably left and right index fingers) and, the vendor submits a set of 5 index finger repeat images captured with the SLI device, from a single finger of a single subject. It is expected that this finger will be completely lifted off of the device and then re-presented to the device for each succeeding scan, i.e., do not simply lay the finger on the device and capture it 5 times in succession. These images shall be submitted for assessment in 8 bpp, monochrome (grayscale) uncompressed form, with polarity such that ridges are dark and valleys are light, in some common format such as TIFF, PGM, BMP, or raw. The images shall not have been previously compressed/decompressed from any lossy compression, such as JPEG or WSQ. If raw format images are supplied, the width and height in pixels, and number of header bytes (if any) must also be supplied. Although BMP format images may be submitted, note that the analysis software on the Test Tools CD (see reference [1] Appendix A) is not compatible with BMP format.
6.3TestProcedures
6.3.1FingerprintGrayRange A single subimage is defined within the image capture area. This subimage is sized and positioned such that it includes most or all of the fingerprint. Specifically, if the background has a constant gray level, then the subimage can be sized to include the entire fingerprint together with an arbitrary amount of background. However, if the background has substantially varying gray levels, then the subimage needs to be sized to avoid the background and only include all, or
SLI3DFingerprintCaptureSpecifications Page32
a substantial portion of the fingerprint. Figure 6.1 illustrates these rules for two cases using a subimage width, height proportional to the image width, height. The gray range is computed within the subimage in each of the 20 finger capture images. This gray range is equal to the total number of gray levels in the subimage which contain signal, where a gray level bin is counted as containing signal if it contains at least a minimum number of pixels. Background: - Since a subimage contains tens of thousands of pixels, the expectation is that if a given gray level bin contains signal, then it would be populated by more than just a few pixels, since all signal pixels are spread out between no more than 256 gray level bins. Therefore, if a gray level bin contains very few pixels it is probably just noise; e.g., dark current, crosstalk, or amplifier noise. A threshold value of 5 pixels can be used to separate gray level bins populated only by noise, from bins populated by signal (+ noise). - The definition of gray range in this section is not necessarily equal to the simple difference between maximum gray level and minimum gray level, e.g., if there is a break/gap in the gray level histogram the two quantities are not the same. - This analysis is performed by the grayfinger software application, which is on the Test Tools CD (see reference [1], Appendix A). In grayfinger, the user selects a subimage size based on a percent of the image width and height, and that subimage is then centered within the total image area.
Figure 6.3: Example subimage Definition for Fingerprint Gray Range Assessment. Image from ref [1].
SLI3DFingerprintCaptureSpecifications Page33
6.3.2FingerprintAbnormalities The fingerprint images are examined to determine the presence of abnormalities which are due to the device with its associated image processing. Initial assessment is performed by close visual inspection of the displayed images. If a significant abnormality is visually detected, then the next stage is to decide whether it is on the input finger (which would be ignored), or is due to the PIV device with its associated image processing. If it is due to the PIV device or image processing, then appropriate image analysis techniques are applied to measure and quantify it11. In some cases, if there is cause for concern that false detail or cosmetic image restoration effects may exist in the fingerprint images, then the FBI may request comparison livescans captured from the same subjects, same fingers as with the PIV device, but using an FBI IAFIS Appendix F certified plain livescan device. Following are some types of abnormalities which may be investigated, among others, depending on the images: - jitter noise effects - localized offsets of fingerprint segments - sensor segmentation / butt joints - noise streaks, erratic pixel response - gray level saturation - false detail - cosmetic image restoration effects In addition, the 5 repeat scans of a single finger are examined for any abnormalities and the degree of reproducibility.
6.3.3FingerprintSharpnessandDetailRendition Qualitative assessment of sharpness and detail rendition is performed by close visual inspection of the displayed fingerprint images, e.g., by comparing a given image to others in the same set, or comparing a given image to other compatible sets of images. Quantitative assessment of sharpness and detail rendition is performed by applying an objective quality metric suitable for fingerprint assessment. In applying such a metric, it is important to keep in mind that the goal is to assess PIV device quality, not input finger quality. Information on attributes, terms, and definitions of biometric image quality metrics and quality scoring is available [IQ]. Although no definitive fingerprint quality metric is identified here, the following lists some metrics which can be useful for fingerprint image assessment. • The NIST Fingerprint Image Quality (NFIQ) metric described in the document: Fingerprint Image Quality, by E.Tabassi, C.L.Wilson, C.I.Watson, NISTIR 7151, August 2004; document and information on obtaining software available at: http://www.itl.nist.gov/iaui/vip/fing/fing.html • The MITRE Image Quality for Fingerprints (IQF) metric.
SLI3DFingerprintCaptureSpecifications Page34
6.4RequirementsCompliance Requirement #1, Fingerprint Gray Range, is met if the gray-scale dynamic range is at least 150 gray levels within the measurement windows in at least 80% of the test fingerprint images. Requirement #2, Fingerprint Abnormalities, is met if: (a) no abnormalities are detected in the fingerprint images or, (b) it is concluded that none of the abnormalities that are detected in the fingerprint images are due to the PIV capture device or its associated image processing or, (c) it is concluded that none of the abnormalities that are detected in the fingerprint images, which are due to the PIV capture device or its associated image processing, are significant enough to adversely impact support to subject authentication via one-to-one fingerprint matching or other intended applications of the PIV device. Requirement #3, Fingerprint Sharpness and Detail Rendition, is met if the sharpness and detail rendition is sufficient to support subject authentication via one-to-one fingerprint matching and other intended applications of the PIV device.
7. ListofReferences
1. Norman B. Nill, “TEST PROCEDURES FOR VERIFYING IMAGE QUALITY REQUIREMENTS FOR PERSONAL IDENTITY VERIFICATION (PIV) SINGLE FINGERPRINT CAPTURE DEVICES,” MITRE Technical Report MTR060170R3, MITRE, Bedford Massachusetts, December 2006.
2. GIF movie examples of Phase Measuring Profilometry (PMP) Structured Light Illumination (SLI) scanning: http://www.engr.uky.edu/~lgh/data/ExamplesofPMP_SLI.htm
3. Principles of Communications, Systems, Modulation, and Noise by R, E. Ziemer and W. H. Tranter, 6th Edition.
4. P. S. Huang, Q. Hu, F. Jin and F-P Chiang, “Color-encoded digital fringe projection technique for high-speed three-dimensional surface contouring,” Optical Engineering, 38(6) pp 1065-1071, (June 1999).
5. Norman B. Nill, “Test Procedures for Verifying IAFIS Image Quality Requirements for Fingerprint Scanners and Printers,” MITRE Technical Report MTR05B0000016, MITRE Bedford Massachusetts, April 2005.
A. APPENDIX:FOURIERCOMPONENTREPRESENTATION The original [1] 2-D CTF and MTF are based on the modulation index which is shown in Subsection A.1 to be dependent on the “dc” or 0 order Fourier [3] component of the captured image. Depth information does not work in the same way since the average surface depth or “dc”
SLI3DFingerprintCaptureSpecifications Page35
can be any absolute depth value and as the DMTF decreases, the measured surface asymptotically converges to the average surface hieght. Furthermore, standard height surfaces may vary in design and not be limited to step edges, 3-D Ronchi rulings or sine wave surfaces. Surfaces made for height standardization already exist for nano-accuracy technologies such as stylus based machines, interferometers and atomic microscopes. Thus, we need a way to define performance measures that are independent of the surface cross-section. Representing a surface cross-section as a set of Fourier Components allows the experimenter to access the different components separately in order to obtain DMTF measures, SNR performance and harmonic (i.e., component) distortion of the surface. The Fourier Transform (FT) of a function g(x) is given by:
dxxfxgfG 2exp (A.1)
Because we have a finite (N) set of sampled values, we approximate the Fourier Transform with the Discrete Fourier Transform (DFT) such that
1
0
2exp][
N
n N
nkngkG
(A.2)
The DFT result is what we use to demonstrate the FT spectrum analysis in the following subsections.
A.1FourierComponentAnalysisofBlurringwith2Dand3Ddata In Fig. A-1, we show the original non-attenuated reflected sine wave superimposed with a blurred, or attenuated sine wave. The device is adjusted manually or with an Automatic Gain Control (AGC) to span between 0 and 1 intensity units. So when blurring occurs, the DC value remains the same but the “AC” sine wave is attenuated. In the original specifications of reference [1], a modulation index of
minimummaximum
minimummaximumIndexModulation
(A.3)
The Modulation Index acts to normalize out the effect of an AGC operation in a typical imaging system. This is probably why it is in subsection 4.4 of reference [1]. The blurring process is represented by a multiplication in the Fourier Frequency Domain as shown in Fig. A-2. The Fourier Transform (FT) of the input 0 to 1 sine wave in Fig. A-1 is shown in Fig. A-2 (a) where the dc component has a value of 0.5 and the two first order components have an amplitude of 0.5 = 0.25 + 0.25. Splitting the non-dc component amplitudes between +f and –f is a characteristic of the FT. A commonly used blurring function is the Gaussian function shown in Fig. A-2 (b). In the frequency domain, this function is multiplied by the input in Fig. A-2 (a) resulting in the same DC level but attenuated sine wave first order components (e.g., harmonics) in Fig. A-2 (c).
SLI3DFingerprintCaptureSpecifications Page36
A.1: Conventional 2‐D contact sine wave pattern attenuation caused by blurring.
In fact, the AGC may change the DC value and scale the blurred sine wave so its maximum is unity. The result of this action would be equivalent to multiplying the Maximum and Minimum by the same coefficient. In the modulation index, this coefficient is scaled away but the result is still dependent on the DC value of the blurred sine wave.
A.2: Fourier domain representation of blurring. (a) Input sine wave spectrum with DC component. (b) Gaussian blurring function. (c) Output spectrum with attenuated first harmonics.
A 3-D point cloud “blurring” process works differently than the 2-D one. Primarily, there is no AGC or manual adjustment of height to change the “DC” level since the DC level represents the average surface height which is an absolute coordinate value. Thus, the average surface height is asymptotically constant and does not need to be part of the calculation. So the average surface depth is subtracted out as shown in Fig. A-3. It is unusual to represent this 3-D process with a Gaussian blurring function but we chose this to make the problem as analogous to the 2-D scenario, as possible. The FT of Fig. A-3 is shown in Fig. A-4. It can be seen in Figs. A-3 and A-4 that the “DC” component is zero but the first order component is the same as in the 2-D case.
SLI3DFingerprintCaptureSpecifications Page37
A.3: 3‐D surface height attenuation (i.e., depth "blurring"). The average surface height is subtracted out.
A.4: Fourier domain representation of depth blurring. (a) Input sine wave surface spectrum with no DC component. (b) Gaussian blurring function. (c) Output spectrum with attenuated first harmonics.
So in the case of 3-D measurement of MTF, there is no need to divide out a scaling coefficient and “DC” is not included in the evaluation of the attenuation. The p-p values of the sine waves are in absolute measurement value such as mm’s. So if the surface really has an ideal sine wave then the p-p value can be measured directly from the surface as in Eq. A.4.
zzpp imumimumfz minmax (A.4)
But, if the surface is not an ideal sine wave then twice the 1st order component amplitude can be used in place of Eq. (A.4).
A.2FourierComponentAnalysisof3DRonchiRulingSurface In Subsection A.1 we discussed the Fourier Components of a sine wave with and without a DC component. Here we consider the FT of a cross section of a Ronchi ruling, also referred to as a square wave. An example is shown in Fig. A-5.
SLI3DFingerprintCaptureSpecifications Page38
The squarewave in Fig. A-5 is filtered in such a way that the output roles off the square edges. It should be noted that some filtering as well as some non-linear processes in 3-D can actually sharpen the edges with “resonant” peaks. But for discussion, we will only consider the “blurring” effect. In Fig. A-6, the FT of the ideal square wave is shown. The tallest peak in the middle is the dc component, which may be set to zero in the case of 3-D data. On either side of dc there are harmonic components. For a square wave, the odd harmonics, 1, 3, 5, … etc. are non-zero while the even harmonics 2, 4, 6, … etc. are zero or “nulls.” Some of these harmonics are indicated on the right side of Fig. A-6. The FT shown is Figs. A-6 and A-7 are approximated by a Discrete Fourier Transform (DFT) so the noise between the harmonics is due to sampling errors but in the ideal case these frequency regions would all be zero. Also, the harmonics occur every 5 cy/mm because the 1st order harmonic is 5 cy/mm and its harmonic frequencies are integer multiples of that.
A.5: Original square wave and filtered square wave.
SLI3DFingerprintCaptureSpecifications Page39
A.6: Harmonics components of a 50% duty cycle square wave with DC.
Any periodic cross section has these harmonics spaced apart by the 1st order harmonic location but they vary in relative amplitude and are not necessary null values. Thus, this approach of selecting out the 1st order component for MTF calculation would work on any cross sectional shape that might be used for a test surface. The “blurred” square wave in Fig. A-7 shows how the high orders tend to be suppressed more than the first order. Therefore, the square wave begins looking more like a sine wave. Another use of the upper harmonic values is the detection of noise. If the system is noisy, the higher frequencies would have noise and the distribution of the noise would indicate what type of noise is present. Noise levels equal at all frequencies would be “white” noise versus low frequency noise would be “pink” noise, etc.
SLI3DFingerprintCaptureSpecifications Page40
A.7: Filtered or attenuated square wave cross section.
SLI3DFingerprintCaptureSpecifications Page41
Figure A.8: Example of a suitable 3‐D composite test target.
B. APPENDIX:FLATTENINGDILEMMA;ORANGEPEELANDSKINSTRETCHDEFORMATION
The “Flattening Dilemma” can be explained by example. Consider a 3-D finger surface that is represented by a cylinder where the cross section is an upper hemi-circle. Using most flattening algorithms, the surface of this half cylinder can be flattened without any map deformation or what we will call “orange peel” deformation. At first impression, a flattening algorithm that can do this would be regarded as desirable. Now consider an algorithm that is designed to flatten a real 3-D fingerprint and the minutia positions are compared with a 2-D contact print and they are shown to be exactly the same. Use this same flattening algorithm to flatten the cylinder model and we find that the surface is laterally distorted. Why would that be? The reason is that flattening a real fingerprint is complicated by not just the surface manifold but also the underlying flesh, bone, pressure of the print and the interaction between the friction ridges and the contact surface. The dilemma is that if you use a simple gridded manifold to test flattening, the target results will converge away from the solution needed to simulate the real flattening of a finger. While this error may be small, it is measurable. For our discussion, we divide deformation into two components, the traditional map deformation or “orange peel” deformation
SLI3DFingerprintCaptureSpecifications Page42
and the “skin stretch” deformation which includes skin compression. The latter deformation is the result of all non-ideal contributions from a real finger. The relationship of the deformation between 3D and 2D methods can be made more clear by a non-statistical experiment with some data our group took in 2008 with a prototype SLI scanner and an OEM contact system. It is non-statistical because we are only going to process a couple fingers on one subject (i.e., we did not have the funding or the objective to conduct a full study of this issue). A set of data demonstrating the deformation on subject 6, digit 3, is shown in Fig. B-1.
Figure B.1: (left) Non‐contact captured image prior to processing. (center) OEM flat fingerprint. (right) Flattened 3D fingerprint. The rectangular regions show where we manually measured the ridge widths.
The left image of Fig. B-1 is the raw camera capture from the SLI device. We refer to this image as the “Mat5 C” image. The center image is the fingerprint obtained from an OEM contact fingerprint device and the right image is the flattened 3D print. We carefully numbered the ridges from the center outward along the red cross section line on all three images, as shown in Fig. B-1. We then selected a center region and two outer regions to manually measure the average ridge width of 4 ridges in each region. As indicated by the tilted rectangles, the measurement was manually made to be about orthogonal to the ridge directions. We conducted these measurements on two fingers of subject 6, digit 3 (shown in Fig. B-1) and digit 5. We also conducted the same measurements on the 3-D data shown in Fig. B-2 where we carefully measured each of the test ridges within 3-D space, “as the crow flies”, to achieve a true non-contact ridge width.
SLI3DFingerprintCaptureSpecifications Page43
Figure B.2: Two views of reference 3‐D data for subject 6, D3 measurements.
To normalize out the units of pixels in Fig. B-1 (left) and inches in center and right and then millimeters in Fig. B-2, we found the ratios between inner ridge widths and outer ridge widths as shown in Table B-1, column 2. Table B‐1: Inner divided by outter (i/o) ratios for digit D3.
Type of Scan Ratio of inner/outer ridge widths i/o ratios normalized by 3D i/o ratio Mat5 C 1.600 1.440 OEM 0.889 0.800 Flattened 3D 1.200 1.080 3D 1.111 1.000 The second column ratios are include both the actual difference between the inner and outer ridge widths as well as the device deformation. The third column of Table B.1 is column 2 divided by the 3-D i/o ratio of 1.111. If we assume the 3D ridge width measurements are the most accurate, then dividing by this ratio removes the variation between the inner and outer ridge width averages leaving ratios that are primarily the deformation caused by the capture or flattening process. If the ratio in column 3 is greater than 1 it indicates that the inner ridge widths were wider than the outer ones due to deformation. Thus, for the mat5 C image, the outer ridges deformation by the surface curvature as expected having the outer ridge widths compressed (1.440). Interestingly, the OEM scanner shows the opposite type of deformation where the inner ridge widths are smaller than the outer (0.800). While this is not a statistical test, let’s look at digit D5 as shown in Table B-2. Digit D5 has different inner to outer ridge ratios but the normalized ratios in column 3 tell a similar story to the results in Table B-1. The bottom line is that there is measurable deformation in both the SLI Mat5 C and the OEM flat print at the camera capture level which means that the sampling is not uniform.
SLI3DFingerprintCaptureSpecifications Page44
Table B‐2 : Inner divided by outter (i/o) ratios for digit D5.
Type of Scan Ratio i/o for D5 i/o ratios normalized by 3D i/o ratio Mat5 C 1.000 1.077 OEM 0.829 0.892 Flattened 3D 0.923 0.994 3D 0.929 1.000 Thus, our argument for keeping the specification for 500 ppi in Section 2 to be as simple for 3-D SLI as it is for 2-D contact fingerprint devices. And our argument for using the direct comparison between minutia locations of a flatten 3-D print versus a 2-D contact be used as the certification of the flattening algorithm plugin.
C. APPENDIX:GRAYLEVELENCODING There are several ways to encode gray level and we limit our discussion to methods that utilize the same data that the depth is calculated from. In this way, there are no alignment issues between the gray level encoding and the depth measurement. There are four methods of determining a texture image. The first one in subsection C.1 corresponds to 3-D certification and is based directly on the depth of the fingerprint curvature. The other 3 are in subsection C.2 and include non-structured illumination, average pattern intensity [4] and pattern intensity modulation.[4]
C.1DepthEncoding is flattened but the ridge depth cross section is preserved. A ridge depth cross section is shown in Fig. C-1. The envelopes for the peaks and the valleys are shown in Fig. C-1.
SLI3DFingerprintCaptureSpecifications Page45
Figure C.1: Gray level coding of the flattened finger surface based on local ridge variation.
The peak envelope is scaled to be the dark gray level and the valley envelope is scaled to be the light gray level. A classic example of this type of depth coding is shown in Fig. C.2 from data collected in 2006 from University of Kentucky prototype scanners. The palmprint scan in Fig. C.2 was performed at about 480 ppi. The top half of the image is the camera image and the bottom half is depth encoded based solely on local depth variation. The fingerprint scan in Fig. C.2 was performed at about 700 ppi. The top half is depth encoded and the bottom half is the camera image.
SLI3DFingerprintCaptureSpecifications Page46
Figure C.2: (Top‐left) 3‐D Palm scan at about 480 ppi. (Top right) Cropped area of top‐left image. (Bottom‐left) 3‐D Scan of Fingerprint at about 700 ppi. (Bottom‐right) Crop of bottom‐left image.
C.2AverageandModulationIntensity. There are three other intensity texturing methods common to SLI methods. All three are related to the 2-Dimensional MTF so they are not used for certification but may certainly be of use I automated as well as human fingerprint matching. The simplest method for obtaining a texture image is to project a solid white pattern and capture the texture image. Ambient and external light can also be used in this way, and if color is desired, a solid red, then green and then blue light is projected and captured to create the RGB color image. Immediately after the texture image is captured, the SLI pattern sequence is projected, captured and reconstructed into a 3-D surface where the points line up with the texture intensities exactly. The average and modulation intensity texture images are obtained from the pattern projection reflected intensities described by
SLI3DFingerprintCaptureSpecifications Page47
zyxN
nyfyxByxAyxI yn ;,22cos,,, C.1
The average intensity is defined as
1
0
,1
,N
nn yxI
NyxA , C.2
and the modulation intensity is obtained from two components defined as
N
nyxIN
yxIN
nnr
2cos,1
,1
0
C.3
and
N
nyxIN
yxIN
nni
2sin,1
,1
0
C.4
Eq. C.3 is referred to as the real component and Eq. C.4 is the imaginary component. The modulation intensity is defined as
yxIyxIyxB ir ,,, 22 C.5
The modulation intensity has useful qualities in part because it eliminates ambient light and reveals an image that corresponds to the SNR quality of the depth measurement.
D. APPENDIX:RECTILINEARCOORDINATES As in reference [1] and [5] we use a rectilinear reference frame for our performance measures. But instead of a 2-D we use a 3-D reference frame. The advantage of rectilinear coordinates is that they are orthogonal and easily measured because their dimensions are straight lines and correspond to the same rectilinear positions of camera pixels. However, in terms of sampling, the characteristics are dependent on orientation of the samples with respect to the dimensions. The sample variation within the X-Y lateral dimensions is also different from sampling along the Z dimension. Consider first just the sampling within the X-Y plane. For example, if a bar chart is aligned with a 2-D camera plane at 500 ppi then we can say in the direction of the pixel rows or columns that the sampling “rate” is 500 ppi. But if we rotate the bar chart by 45 degrees the pixel spacing in that direction increases by 1.414 (i.e., the square root of 2) and the sampling rate appears to decrease to 353.6 ppi = 500 ppi/1.414. However, it is more complicated than this because at 45 degrees, the diagonal pixels at 353.6 ppi are alternately shifted in space. This complicates the actual sampled image and the combined effect of the closer
SLI3DFingerprintCaptureSpecifications Page48
diagonal rows actually increases the average sampling rate. In the extreme, the adjacent diagonals bias the sampling from 353 ppi toward 707 ppi = 2 x 353.6 ppi. To demonstrate this, we took a bar chart and placed a pin in the middle from which we rotate the chart. Then with the bars vertical and at about a 45 degree angle we captured the images. We purposely oversampled the image with a high resolution camera so we could systematically downsample both images equally until some interference patterns began to form.
Figure D.1: (left) Vertical bars showing aliasing with interference patterns. (right) The same resolution, the 45 degree angle chart shows little aliasing.
Fig. D.1 (left) is showing indications (i.e., variations of line periods) of interference patterns forming which are related directly to aliasing. The line widths in Fig. D.1 (right) are much more uniform and the contrast is also higher and more uniform, even though the sampling along the diagonal is significantly less. This is due to the diagonal spacing of the lines of sample points are actually closer. The geometry of this angular sampling effect is shown in Fig. D.2. Fig. D.2 (left) corresponds to the sampling geometry of Fig. D.1 (left) and Fig. D.2 (right) corresponds to Fig. D.1 (right). This increase in sampling rate at 45 degrees becomes more apparent when considering the pixel footprints as shown in Fig. D.2. Since the pixels effectively integrate adjacent regions and any optical blurring also integrates spatial areas then a diagonal line will effectively intersect a footprint 707 times per inch if the rectilinear spacing is 500 ppi.
Sampling do to out-of-plane rotation or the depth surface gradient has different characteristics.
SLI3DFingerprintCaptureSpecifications Page49
Figure D.2: (left) Sample spacing correspondence with camera pixels aligned with the bar chart patterns. (right) Sample spacing with Camera pixels aligned at 45 degrees to bar chart pattern.
Figure D.3: (left) Cross section of fingerprint. (right) Cropped region showing gradient effects on sample spacing.
Fig. D.3 shows the cross section of a 3-D fingerprint capture prior to flattening. Fig. D.3 (right) is a cropped section with dashed lines intersecting the cross section at different gradients. It can be seen that the effective sample spacing after flattening would be non-uniform. Sample spacing in the cross sectional direction of a ridge is dependent on the ridge cross section more than the cross section overall curvature. In the orthogonal direction or along a ridge, the sample spacing is dependent of the overall curvature. This type of non-uniformity is more complicated than a contact flattened print. The contact plane print is effectively non-uniformly sampled but because the ridges would be flattened, but the sample spacing is much less dependent on the overall curvature of the fingerprint.
F. APPENDIX:APPENDIXF,PIVandSLICROSSREFERENCING
SLI3DFingerprintCaptureSpecifications Page50
For multi-finger and palm print scanner certification, reference [5] is commonly referred to as “Appendix F” certification. The PIV [1] is specific to 500 ppi single fingerprint scanner certification while the Appendix F [5] includes 1000ppi, slightly different levels of performance, and extends the certification to multi-finger and palm scanners. However, as far as the type of testing, the Appendix F, PIV and our SLI have similar tests. So in Table F.1 we provide a cross reference between the 3 documents that index to where each test is, the value changes necessary and additional or different test types. The acronyms used for the documents are Appendix F (ApF) for reference [5], PIV for reference [1] and SLI for this specification. Table F‐1: Appendix F, PIV and SLI Specification Cross Reference Table
Appendix F (ApF) [5] PIV [1] SLI 2.1 Linearity
PIV: not in PIV. But noted in section 4.6 SLI: See Section 3.7 and 5.4
