Note: Learningwithout Human Scores for Blind Image Quality Assessment

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