BLeSS:
Bio-inspired Low-level Spatiochroamtic Similarity Assisted Image Quality Assessment
School of Electrical and Computer Engineering, Center for Signal and Information Processing
DoganCan Temel and Ghassan AlRegib
Georgia Institute of Technology, USA
CENTER SURROUND EFFECTS
MOTIVATION
CONTACT US
[email protected] [email protected] cantemel.com ece.gatech.edu/research/labs/msl/
E M A I L W E B
Application Average daily
shared photos
390 Million
700 Million
70 Million
760 Million
[1]
Billions of images are shared every day.
Supported resolution: not very high
Supported resolution by display systems:
up to ultra HD
Applications started to catch up with
high resolution.
Users are concerned about network usage
To balance quality and network usage, image
optimization requires image quality assessmentRelated Applications
Smart Capture Remote Assistance
PROBLEM DESCRIPTION
Reference images
Distorted images
Subjective ScoresBad1
very
annoying
Poor2 annoying
Fair3slightly
annoying
Good4distortion but not
annoying
Very
Good5no perceived
distortion
Test setup Stimuli
Problem Model
QUALITY
ESTIMATORSFidelity Structure
Scale
Space
Visual
SystemPooling Color
Before
2000
MSE-PSNR
PSNRc Pixel-wise
2000 NQM
2003 MS-SSIM
2004 SSIM
2006 PSNR-HVS
2007PSNR-HVS-M
VSNR
2008VIF
C4 Local Mean
2009CW-SSIM
Li-Wang
2011
PSNR-HA
PSNR-HMA
FSIM
FSIMc Pixel-wise
IW-SSIM
LBIQ
DIIVINE
2012
CIEDE Local Mean
BRISQUE
CORNIA
MLIQM
SR-SIM
2013
CB/SF
QAC
SPARQ
2014
Tang
QAF
Kang
𝑄𝑎𝑟𝑒𝑎
𝑄𝑒𝑥𝑝𝑜𝑛𝑒𝑛𝑡
2015
REDLOG
IQA-CNN++
Li
𝑆2𝐹2
DLIQA
Gao
2016
CNN-SVR
BLeSSCenter
Surround
LITERATURE
SSF is high stimuli ~ surround
SSF is low stimuli !~ surround
Surround Spatial Frequency (SSF)
Surround Orientation
BLeSS PIPELINE
VALIDATIONVISUALIZATION
Distorted ImageReference Image
SR-SIM Map FSIM Map BLeSS Maps
Percentage performance (Spearman correlation)
changes for BLeSS-assisted IQA methods over
distortion categories.
Percentage performance (Spearman correlation)
changes for BLeSS-assisted IQA methods over
full databases.
Spatiochromatic Grouping Pipeline
BLeSS Pipeline
Murray 2013
Murray 2013
SR-SIM FSIM FSIMc
Comp. -0.29 (000) +0.13 (000) +0.28 (000)
Noise -2.16 (001) -1.31 (000) -0.34 (000)
Comm. +0.07 (0-0) +0.25 (0-0) +0.24 (0-0)
Blur -0.39 (000) +0.20 (000) +0.40 (000)
Color +183 (--1) +185 (--1) +13.1 (--1)
Global -1.31 (--0) -4.69 (--0) -0.16 (--0)
Local -1.85 (--0) +4.36 (--0) +3.12 (--0)
SR-SIM FSIM FSIMc
LIVE +0.13 (0) +0.17 (0) +0.06 (0)
MULTI -0.33 (0) +0.62 (0) +0.79 (0)
TID13 +3.79 (1) +4.77 (1) +1.03 (0)
LIVE MULTI TID13 Total
Comp. 460 180 375 1015
Noise 174 180 1375 1729
Comm. 174 - 250 424
Blur 174 315 250 739
Color - - 375 375
Global - - 250 250
Local - - 250 250
The number of distorted images with respect to
degradation categories in each database.
Wavelet
Transform
Grouplet
Transform
Center
Contrast
Normalization
Contrast
Sensitivity
Adjustment
Bicubic
Interpolation
Inverse
Wavelet
Transform𝜏𝑖
Opponent
Color Channel
Separation
Euclidean
Norm𝑖 ∈ 1, 2, 3𝜏𝐼𝑖
𝑖 ∈ 1, 2, 3𝐼
Spatiochromatic
Grouping
Pixel-wise
Similarity
𝐼
Spatiochromatic
Grouping𝐼̃
Mean
PoolingBLeSS
LIVE
LIVE
𝜏
𝜏̃