+ All Categories
Home > Documents > ZhengHuiHu_presentation031907

ZhengHuiHu_presentation031907

Date post: 05-Apr-2018
Category:
Upload: michael-erick
View: 217 times
Download: 0 times
Share this document with a friend

of 35

Transcript
  • 8/2/2019 ZhengHuiHu_presentation031907

    1/35

    California Car License Plate

    Recognition System

    ZhengHui Hu

    Advisor: Dr. Kang

  • 8/2/2019 ZhengHuiHu_presentation031907

    2/35

    Introduction

    A License Plate Recognition System

    (LPRS) is a system to automatically

    detect, recognize and identify avehicle plate.

    It involves low-level image

    processing techniques with higherlevel artificial intelligence

    techniques.

  • 8/2/2019 ZhengHuiHu_presentation031907

    3/35

    Applications

    Mainly for monitoring, surveillanceand security. For example,

    Entrance/Exit monitoring for parking lotstructures

    Part of surveillance system for gatedcommunities

    Control gateways for vehicle passage

    Security Systems for high traffic

    Law Enforcement

  • 8/2/2019 ZhengHuiHu_presentation031907

    4/35

    Challenges

    Image Capturing Vehicle speed

    Lighting condition

    Occlusion

    Processing speed

    Heavy trafficRecognition accuracy

    High correctness

  • 8/2/2019 ZhengHuiHu_presentation031907

    5/35

    Recognition Stages

    Plate Localization Locate plate region out of

    car and/or background

    Character Segmentation Segment each

    character/number out ofplate

    Character Recognition Recognize each character

    on the plate

    Similar to OCR process

    Plate Localization

    Character

    Segmentation

    Character Recognition

  • 8/2/2019 ZhengHuiHu_presentation031907

    6/35

    Previous Work

    Many difference solutions have

    already been proposed for each

    stage of recognitionPlate localization

    Use edge statistics to locate the plate

    Fuzzy clustering algorithms

  • 8/2/2019 ZhengHuiHu_presentation031907

    7/35

    Previous Work

    Character Segmentation Vertical/horizontal projection Adaptive Clustering

    Optical Character Recognition Template matching

    Neural network

    Feature analysis

  • 8/2/2019 ZhengHuiHu_presentation031907

    8/35

    Approach

    Target: California State License

    Plates The word California appears at the

    top-center on the plate in red and italic. The plate number starts with a digit(0-9),

    followed by three English characters (A-

    Z) and three more digits (0-9). The plate background is a light shade of

    gray while its characters are of dark

    blueish color.

  • 8/2/2019 ZhengHuiHu_presentation031907

    9/35

    Input Images

    Captured using adigital camera Different distance

    Different lightingconditions

    Different angles

    Original size2048X1536

    Resized to 800X600for faster process

  • 8/2/2019 ZhengHuiHu_presentation031907

    10/35

    Base Knowledge

    No assumption on size nor the

    possible location.

    Helpful knowledge light background color and dark

    foreground

    rectangular shape with same width andheight proportion high in edge concentration

  • 8/2/2019 ZhengHuiHu_presentation031907

    11/35

    Plate Localization

    Noise Filtering andBrightness normalization

    Extract Edge Information

    Filter using Colorand Edge Information

    Connected Component

    Analysis

    Candidate(s) Found

    Continue to

    Character Segmentation

    > Filter thresholdFailed

  • 8/2/2019 ZhengHuiHu_presentation031907

    12/35

  • 8/2/2019 ZhengHuiHu_presentation031907

    13/35

    Extract Edge Information

    Create a difference

    image using

    equationimg =threshold(close(src)-src)

    High frequency

    features areenhanced by the

    operation

  • 8/2/2019 ZhengHuiHu_presentation031907

    14/35

    Filtering

    Select plate

    background color

    using two criteria Next to edge pixels

    detected in

    previous step

    A light shade ofgray

  • 8/2/2019 ZhengHuiHu_presentation031907

    15/35

    Connected Component

    AnalysisCreate connected

    components

    Find candidateregions

    High concentration

    of edge pixels

    Width/Height ratiosimilar to plate

  • 8/2/2019 ZhengHuiHu_presentation031907

    16/35

    Candidate Regions

    Multiple candidate

    regions can be

    found at this stage

    All of them will be

    submitted to next

    stage

  • 8/2/2019 ZhengHuiHu_presentation031907

    17/35

    Character Segmentation

    Filter and Threshold

    Find componentsof similar size

    regionunprocessed ?

    No

    Count Components

    Resizecomponents

    7

    Interpolate boxes

    < 2> 2 and < 7

    Crop Picture

    Continue to Neural NetworkRecognition

    Failed

  • 8/2/2019 ZhengHuiHu_presentation031907

    18/35

    Filter and Threshold

    Color filter the plate to remove foreignelements

    Apply inverse binary threshold using thefollowing function

    fimgij=255, if imgijthreshold

    0, otherwise

  • 8/2/2019 ZhengHuiHu_presentation031907

    19/35

    Character Extraction

    Find a series of boxes with similar shapeand size. These are the individual

    characters If number of boxes found is less than 7then an interpolation is performed

  • 8/2/2019 ZhengHuiHu_presentation031907

    20/35

    Character Recognition

    Two neural networks are used. One for recognition of digits (0-9) One for characters (A-Z)

    Both networks were trained by using ahybrid method combining traditionalBack-propagation algorithm with a

    Simulated Annealing process

  • 8/2/2019 ZhengHuiHu_presentation031907

    21/35

    Artificial Neural Network

    Artificial NeuralNetworks (ANN)

    are modeled afterthe human brain Network of

    processing units

    called Neurons Good for solving

    classificationproblems

    InputLayer

    HiddenLayer

    OutputLayer

  • 8/2/2019 ZhengHuiHu_presentation031907

    22/35

    Neurons

    Neurons process information byreceiving and firing signal according to

    internal function Two different types are used Threshold based step function

    Sigmoid function

    f=1, if wijthreshold0, otherwise

    f=1

    1et

  • 8/2/2019 ZhengHuiHu_presentation031907

    23/35

    ANN Training

    ANN must be trained by example beforeuse.

    Supervised Training ANN receive set of Input data, and output ANN is adjusted according to the error

    produced Repeat with different set of data

    Back propagation (B-P) algorithm is aclassical supervised training method

  • 8/2/2019 ZhengHuiHu_presentation031907

    24/35

    Limitation of B-P training

    Same limitation ofany gradient

    descent algorithm Lengthy flatplateau travel

    Local minimatrap

    Localmaximum

    GlobalOptimum

    FlatPlateau

  • 8/2/2019 ZhengHuiHu_presentation031907

    25/35

    Simulated Annealing (SA)

    Probabilistic heuristic for locating globaloptimum in large search space Invented

    by Kirkpatrick, S., Gelatt, C.D., andVecchi, M.P. in 1983 Inspired by metallurgic annealing

    process in which metal is cool down

    gradually to get the best configuration Inner random selection allows it to

    escape from local minima trapping

  • 8/2/2019 ZhengHuiHu_presentation031907

    26/35

    Hybrid B-P SA Training

    Train neighbors

    Find neighbors

    Create new ANN, astraining candidate

    Train candidate

    NN with least

    error is new candidate

    reduce T

    End Training

    < thresholdOR T > 0

  • 8/2/2019 ZhengHuiHu_presentation031907

    27/35

    Experimental Results

    Hardware used

    Pentium 4 1.0 GHz processor and 512

    MB of RAMImages

    50 Images were acquired using a digital

    cameraOriginal image size: 2048X1536

    Test image size: 800X600

  • 8/2/2019 ZhengHuiHu_presentation031907

    28/35

    Experimental Results

    Plate Localization

    Character Segmentation

    Total Images Plate located Failed to locate Success rate

    50 48 2 96%

    Total Images Character

    segmented

    Failed Success rate Cumulative success

    rate

    48 45 3 93.75% 90%

  • 8/2/2019 ZhengHuiHu_presentation031907

    29/35

    Experimental Results

    Optical Character Recognition Training 5 images as basis for each digit Create 9 variations by adding noise, altering

    column/rows, distortion, etc.. Recognition rate of digits

    Recognition rate on training data 80%

    Recognition on actual images extracted from LPRS 60%

  • 8/2/2019 ZhengHuiHu_presentation031907

    30/35

    Experiment Results

    Processing time

    Average processing time from image

    input to result: ~ 300ms

  • 8/2/2019 ZhengHuiHu_presentation031907

    31/35

    Conclusion

    Preprocessing

    Proposed algorithm's performance are

    satisfactoryNeural network training

    New method combining Back-

    Propagation algorithm with SimulatedAnnealing process

  • 8/2/2019 ZhengHuiHu_presentation031907

    32/35

    Future Work

    Extend current system to alsorecognize uncharacteristic plates andadditional character set.

    Improve recognition ratio by usingalternative ANN configurations

  • 8/2/2019 ZhengHuiHu_presentation031907

    33/35

    References

    1.William K. Pratt, Digital Image Processing, Third Edition, JohnWiley & Sons, 20012.R. Parisi, E.D.Di Claudio, G.Lucarelli, and G. Orlandi, Car PlateRecognition by neural networks and image processing, Proceedingsof the 1998 IEEE International Symposium on Circuits and Systems,

    (ISCAS '98).3.Leonard G. C. Hamey, Colin Priest, Automatic Number PlateRecognition for Australian Coditions, Proceedings of the DigitalImaging Computing: Techniques and Applications (DICTA 2005)4.Bai Hongliang, Liu Changping, A Hybrid License Plate ExtractionMethod Based on Edge Statistics and Morphology, Proceedings of the

    17th

    International Conference on Pattern Recognition(ICPR'04)5.Choudhury A. Rahman, Wael Badawy, Ahmad Radmanesh, A RealTime Vehicle's License Plate Recognition System, Proceedings of theIEEE Conference on Advanced Video and Signal Based Surveillance(AVSS'03)

  • 8/2/2019 ZhengHuiHu_presentation031907

    34/35

    References

    6.Takashi Naito, TOshihiko Tsukada, Keiichi Yamada, Kazuhiro Kozuka,Shin Yamamoto, Robust License-Plate Recognition Method forPassing Vehicles Under Outside Environment, IEEE Transactions onVehicular Technology, 20007.Rodolfo Zunino, Stefano Rovetta, Vector Quantization for License

    Plate Location and Image coding, IEEE transactions on IndustrialElectronics, Vol 47, No. 1, 20008.Mi-Ae Ko, Young-Mo Kim, License Plate Surveillance System UsingWeighted Template Matching , Proceedings of the 32ndAppliedImagery Pattern Recognition Workshop (AIPR' 03)9.Feng Yang, Zheng Ma, Vehicle License Plate Location Based on

    Histogram and Mathematical Morphology, Proceedings of the FourthIEEE Workshop on Automatic Identification Advanced Technologies(AutoID'05)10.Shyang-Lih Chang, Li-Shien Chen, Yun-Chung Chung, Sei-WanChen, Automatic License Plate Recognition, IEEE transactions onIntelligent Transportation Systems, Vol 5, No. 1, 2004

  • 8/2/2019 ZhengHuiHu_presentation031907

    35/35

    References

    11.Juntanasub, R., Sureerattanan, N., Car license plate recognitionthrough Hausdorff distance technique, Proceedings of the 17th IEEEInternational Conference on Tools with Artificial Intelligence, (ICTAI' 05)12.Timothy Masters, Advanced Algorithms for neural networks?, JohnWiley & Sons, 1995

    13.Cornelius T. Leondes (Editor), ?Algorithms and Architectures(Neural Networks Systems Techniques and Applications), AcademicPress, 199814.David E. Rumelhart, Geoffrey E. Hinton & Ronald J. Williams,Learning representations by back-propagating errors, Nature 323, 533- 536 (09 October 1986)

    15.Kirkpatrick, S., Gelatt, C.D., and Vecchi, M.P., Optimization bySimulated Annealing, Science, Volume 220, Number 4598, 13 May1983, pp. 671680.