Classical measures Mean squared error (MSE): Peak Signal to Noise ratio (PSNR): Original pixel values are xi and distorted xi. Maximum pixel value is 255.
Three quality factors Contrast errors: Structural errors: Quantization errors:
Weighting the factors Weighted sum: Scaling function: Scales the values approximately to the range 0, 1. Negative values are rounded to zero.
Test images Lena (512 512) Airplane (512 512) Bridge (256 256)
Test arrangements Image samples: Offset quality photographs Viewing conditions: Normal office environment Viewing distance: Free of choice for the observer Time of evaluation: Unlimited Number of observers: 15 -39 Grading: Extended MOS scale (0 -10).
Visual examples JPEG (1. 0) BTC QT
JPEG Contrast: Structural: Quantization: BTC QT
Scatter of the quality grades
Quality grades
Correlation to subjective quality
Conclusions • Blockwise measure instead of pixelwise like MSE or PSNR • Correlates better to subjective quality • Old method but worth to study further: – Extension to Multi-scale resolution – Study with newer image datasets – Compare to SSIM 1 and RF 2 1 W. Zhou, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli Image quality assesment: from error visibility to structure similarity IEEE Trans. on Image Processing, 13 (4), 600 -612, 2004. 2 A. Mahmoudi-Aznaveh, A. Mansouri, F. Torkamani-Azar, and M. Eslami Image Quality Measurement Besides Distortion Type Classifying Optical Review, 16 (1), 30 -34, 2009.