基于机器学习的图像质量盲评价

Image Quality Assessment (IQA)FidelityThe distortion level between the test and reference images in the process of observation of the h f h human eye.
MLA2013
IntelligibilityThe ability of the test image on providing information for human / machine consumers with respect to th f t the reference image. i
VIPS Lab, Xidian University

IQA and Its ApplicationsVideo Blogging Printing
MLA2013
Enhancement
Restoration
Monitor image quality in quality control systemsTele-Health
Authentication
Watermarking
Image Acquisition
Channel Coding
IQA
Computer Graphics
Benchmark i imageprocessing systems and algorithmsGaming
Broadcasting
Compression Content Delivery
Retrieval
Optimize the systems and parameter settingsQoS Monitoring
Digital Camera Surveillance
VIPS Lab, Xidian University

Very Satisfied R User satisfaction MOS 100Very Satisfied Satisfied Some Users Dissatisfied 4.34.03.69080Many Users Dissatisfied Nearly All Users Dissatisfied 3.12.61070

6050Not Recommended 1.00

Luminance Shift

(MSE=309)Original Image Contrast Stretch (MSE=306)Impulsive Noise (MSE=313)

JPEG Compression (MSE=309)Gaussian Noise

(MSE=309)Blurring (MSE=308)Spatial Shift (MSE=590)

0.30.350.250.30.351

2

3

456

7

8

-2.5-2

-1.5

-1-0.500.5subband

e

1

2

3

456

7

8

-13-12-11

-10

-9-8-7-6subband

-5

5

00.050.10.150.20.251

2

3

456

7

8

-2.5-2-1.5

-1-0.500.5subband

e

1

23

45678

-13-12

-11

-10

-9-8-7-6subband

-5

5

00.050.10.150.2Features Features

Distortions

Regression Dictionary

Learning g

Representation

Learning

…

Deep

Learning

Features

Marginal Distribution # 2230

Extract

Joint

Distribution

# 14400

PCA

Image Quality

Blur/noise

Statistics

# 57

Compact Features

Extract features that measure aspects of image structure and statistics R d d f f d d i

Redundancy of features are reduced via PCA

Distortion

z Develop an effective BIQA without human scored images for training groups Patch Extraction z

Group patches into different groups,and QAC is applied to each group to learn the quality-aware centroids Patch Extraction Compare each patch with the centroids and

Feature Extraction

Patch Quality

Estimation Mapping Image Quality

Quality Aware

Clustering

(QAC)Function

0.30.350.250.30.351

2

3

456

7

8

-2.5-2

-1.5

-1-0.500.5subband

e

1

2

3

456

7

8

-13-12-11

-10

-9-8-7-6subband

-5

5

00.050.10.150.20.251

2

3

456

7

8

-2.5-2-1.5

-1-0.500.5subband

e

1

23

45678

-13-12

-11

-10

-9-8-7-6subband

-5

5

00.050.10.150.2Features Features

Distortions

?Decision trees

Learning

?Boosting

?Bayesian learning

?Support vector machine

VIPS Lab, Xidian University

SVM

Image features

JPEG

JP2k

Fast

White

Noise

Gblur

Wavelet coefficient

Test image

s

SVM

Quality class 1 Quality class 2Class-1 IQA

Class-2 IQA

Quality class 4Quality class 3Quality class 5Original image Class-3 IQA

Class-4 IQA

Class-5 IQA

SVM Image features Classifier Image quality

Test image Specific IQA Spatial criteria Spatial frequency criteria

Spatial criteria Spatial-frequency criteria

0.30.350.250.30.351

2

3

456

7

8

-2.5-2

-1.5

-1-0.500.5subband

e

1

2

3

456

7

8

-13-12-11

-10

-9-8-7-6subband

-5

5

00.050.10.150.20.251

2

3

456

7

8

-2.5-2-1.5

-1-0.500.5subband

e

1

23

45678

-13-12

-11

-10

-9-8-7-6subband

-5

5

00.050.10.150.2Features Features

Distortions

Support

Vector Machine

Features Image quality

Image Neutral

Network

Image quality Image

本文来源：https://www.bwwdw.com/article/gcnq.html