Evaluation of Two Principal Image Quality Assessment Models

1
Evaluation of Two Principal Image Quality
Assessment Models

• Martin Cadík, Pavel Slavík
• Czech Technical University in Prague, Czech
  Republic
• cadikm_at_sgi.felk.cvut.cz

2
Content

• Image Quality Assessment
• Traditional error sensitivity approach, VDP
• Structure similarity approach, SSIM
• Traditional vs. Structural Approach
• Conclusion

3
Image Quality Assessment

• Assessing the quality of images
• image compression
• transmission of images
• Subjective testing
• the proper solution
• expensive
• time demanding
• impossible embedding into algorithms

4
Image Quality Assessment Models

• RMSE is NOT sufficient

MODEL( , )
Detection probability map
5
Image Quality Assessment Computer Graphics

• Quality improvement
• Saving of resources
• Effective visualization of information
• etc.

6
Error Sensitivity Based Approach

• General framework
• Visible Differences Predictor Daly93
• Perceptual Distortion Measure Teo, Heeger 94
• Visual Discrimination Model Lubin 95
• Gabor pyramid model Taylor et al. 97
• WVDP Bradley 99

7
Visible Differences Predictor

• Daly 93
• Threshold sensitivity
• Visual Masking

8
Structural Similarity Based Approach

• Main function of the HVS to extract
  structural information
• UQI Wang 02
• SSIM Wang 04
• Multidimensional Quality Measure Using SVD
  Shnayderman 04

9
Structural SIMilarity Index

• Wang 04
• Simple implementation
• Fast computation

10
Traditional vs. Structural Subjective Testing

• Independent subjective tests
• 32 subjects
• 30 uniformly compressed images (JPEG2000)
• 30 ROI compressed images
• difference expressed by ratings
• Mean Opinion Scores

11
Traditional vs. Structural Objective Testing
Original (left) and ROI compressed (right) input
images SSIM probability map (left) and VDP
probability map (right)
12
Traditional vs. Structural Test Results
Quality predictions compared to subjective MOS
for the SSIM (left) and for the VDP (right)
13
Traditional vs. Structural Test Results (cont.)
Quality assessment performances of the SSIM and
for the VDP models CC Pearson (parametric)
correlation coefficient SROCC Spearman
(non-parametric) correlation coefficient
14
Conclusion

• Independent comparison of two IQA approaches
• VDP, SSIM
• subjective data (uniform/ROI)
• Results
• SSIM better
• SSIM faster to compute and easier to implement
• both models perform badly in ROI tasks
• SSIM can detect the ROI
• gt SSIM significant alternative to thoroughly
  verified VDP

15
Thank You for Your Attention

• ANY QUESTIONS? cadikm_at_sgi.felk.cvut.cz
• ACKNOWLEDGEMENTSThis project has been partly
  supported by the Ministry of Education, Youth and
  Sports of the Czech Republic under research
  program No. Y04/98 212300014, and by the CTU in
  Prague - grant No. CTU0408813. Thanks to Radek
  Vaclavik and Martin Klima for their support
  during the subjective testing.