Xue W, Zhang L, Mou X. Learning without human scores for blindimage quality assessment[C]//Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition. 2013: 995-1002.
1. Abstrct
a) Shortcomingsof exsiting methods in BIQA
i. Needa large number of training images
ii. Lackan explict explanation
iii. Usinghuman scored images
b) Proposea novel approaches
i. Partitionthe distorted images into overlapped patches
ii. Usea percentile pooling strategy to estimate thelocal quality
iii. Proposea quality-aware clustering(QAC) method to learn a set of centroids
iv. Usethe centroids as codebook to infer the quality of each patch
v. Obtaina perceptual quality score of the whole image
c) Meritsof QAC
i. Simpleand effective
ii. Comparableaccuracy to those methods using human scored images
iii. Highlinearity
iv. Real-timeimplementation
v. Availabilityof image local quality map
2. Introduction
a) Reasons of images beingdeteriorated
i. Noise corruption
ii. Blur
iii. JPEG compression
iv. JPEG 2000 compression
b) Two categoriesof BIQA
i. Distortion specific methods:quantify the particular artifacts
ii. Distortion independent methods: generalperpose BIQA,lack of distortion information
c) Databases
i. TID2008
1. 25 reference images and theirdistorted versions of 17 types on 4 levels
ii. LIVE
1. 779 distorted images generatedfrom 29 original images
2. 5 types of distortions:JPEG2000\JPEG\additive white noise\Gaussian blurring\simulated fast fadingRayleigh channel
iii. CSIQ
1. 30 original images and theirdistorted images
2. 6 types of distortions on fivedifferent distortion levers
d) Learning methods with humansubjective quality scores
i. Two-step framework:
1. Feature extraction
2. Model regression
ii. Moorthy:SVM(to detect thedistortion type) +SVR
iii. Saad:probabilistic model basedon the contrast and structural features
iv. SVR+ radial basis function
v. Sparse representation
e) Several important issues above:
i. Need a large amount of humanscored images for training
ii. A black box and relationshipbetween features and quality score implicit
iii. Complexity id too high
f) Mittal:pLSA(probabilisticlatent semantic analysis)
g) Propose a novel method based onQAC
i. Partitionthe distorted images into overlapped patches
ii. Usea percentile pooling strategy to estimate thelocal quality
iii. Proposea quality-aware clustering(QAC) method to learn a set of centroids
iv. Usethe centroids as codebook to infer the quality of each patch
v. Obtaina perceptual quality score of the whole image
h) Meritsof QAC
i. Simpleand effective
ii. Comparableaccuracy to those methods using human scored images
iii. Highlinearity
iv. Real-timeimplementation
v. Availabilityof image local quality map
i) Organization
i. Section2: describe the learning of quality-aware centroids by QAC
ii. Section3: use the learned centroids to perform blind quality estimation
iii. Section4: expriments and discussions
iv. Section5: conclusion
3. Quality-awareclustering
a) Learningdataset generation
i. Randomlyselect from the Berkeley image database
ii. Simulatethe distorted images of the ten images
1. Gaussiannoise
2. Gaussianblur
3. JPEGcompression
4. JPEG2000compression
iii. Threequality levels
1. noisestandard deviation(Gaunssian noise)
2. Blurkernel(Gaussian blur)
3. Resultedquality level(JPEG)
4. Compressionratio(JPEG2000)
iv. 120distorted images and 10 reference images
b) Patchquality estimation and normalization
i. Assigna perceptual quality:SSIM,FSIM
ii. Normalizatein order to make the average of all si in an image as close to its overallperceptual quality as possible
c) Quality-awareclustering
i. Group intogroups of similar quality
ii. Clusterthose patches in the same quality group into different clusters based on theirlocal structure
iii. Highpass filter,the output of on thethree scales are concatenated into a feature vector
iv. K-meanclustering algorithm
v. HaveL sets of centroids on L different quality levels,call them quality-awarecentroids.
4. Blindquality pooling
a) Patchpartition and feature extraction
i. Partitioninto N overlapped patches
ii. thenuse high pass filter to extract the feature vector
b) Clusterassignment
i. Wefind the nearest centroid to the feature ofpatch
ii. Weassign to Lclusters definded by
c) Patchquality score estimation
i. Qualityscore of
d) Finalpooling
i. Simplestaverage pooling
5. Exprimentalresults
a) Protocol
i. Threelargest publicly available subject-rated databases
1. LIVE
2. CSIQ
3. TID2008
ii. Subjectivequality score
1. Meanopinion score(MOS)
2. Differencemean opinion score(DMOS)
iii. Commontypes of distortions
1. JPEG2000compression
2. JPEGcompression
3. WN
4. GB
iv. Correlationcoefficients between the prediction results and the subjective scores
1. Spearmanrank order correlation coefficient(SROCC)
2. Pearsoncorrelation coefficient(PCC)
b) Implementationdetails and results of QAC
c) Comparisonwith state-of-arts
d) Computationalcomplexity
6. Conclusions