180 likes | 329 Views
A Comparison of Quality Metrics for JPEG Images. Feng Xiao Fall 2000. Motivation. Compare performance of different image metrics for JPEG images with subjective measurement Blocking is the dominant artifact in JPEG images (or other block-based coding), especially at low-bit-rate
E N D
A Comparison of Quality Metrics for JPEG Images Feng Xiao Fall 2000 EE368B
Motivation • Compare performance of different image metrics for JPEG images with subjective measurement • Blocking is the dominant artifact in JPEG images (or other block-based coding), especially at low-bit-rate • Post-processing may incur blurring when reducing blocking • Need a good metrics EE368B
Candidate Metrics • RMSE (root-mean-square error) • BMR (block-to-mask ratio, Liu 1997) • EOBD (effect-of-block-distortion, Eskicioglu 1995) • MIX (RMSE + BMR) • RMSE is pixel-based, and BMR is block-based, combination may be more robust EE368B
BMR: I • Compute the block difference 6 7 8 9 1 Block Border 2 EE368B
BMR: II • Include the perceptual effects where is the just-noticeable difference 50 is a weighted ratio EE368B
BMR: III • Separate the blocking and blurring measure • OBMR(i,j): BMR in the original image • PBMR(i,j): BMR in the processed image. • a) PBMR(i,j) > OBMR(i,j). Block(i,j) in processed image is more blocking than that of the original image. • b) PBMR(i,j) <= OBMR(i,j). Block(i,j) is blurred in processed image. • blocking strength = mean(|OBMR(i,j)-PBMR(i,j)|) for set a • blurring strength = mean(|OBMR(i,j)-PBMR(i,j)|) for set b EE368B
BMR: IV • Construct the single BMR BMR= blocking strength + blurring strength EE368B
BMR: V Strength Strength JPEG quality Size of smoothing filter EE368B
EOBD EE368B
Experiments Click on the image with the worst quality JPEG with de-block JPEG JPEG with Filtering (3x3) EE368B
Experiments (cont.) • Each experiment has18x3 images: • 18 JPEG images at quality levels 5~40 (bits .25~.80 bpp) • 18 smoothed (3x3) JPEG images • 18 de-blocked JPEG images (Chou’s 1995) • Repeat 4 times • 2 subjects, 2 image sets (‘lena’ & ‘einstein’) EE368B
Mean Rank Error RMSE BMR MIX EOBD Results: Comparison Rank Error for Image i: Ei= | Si – Ri |, where Si is the subjective rank of image I, Ri is the rank derived from metrics EE368B
Results: Post-processing Improvement (rank order) Bit Rate (bpp) EE368B
Results: RMSE vs. Subjective RMSE Subjective Rank Order EE368B
Results: BMR vs. Subjective BMR Subjective Rank Order EE368B
Results: EOBD vs. Subjective EOBD Subjective Rank Order EE368B
Results: MIX vs. Subjective MIX Subjective Rank Order EE368B
Conclusion • MIX is the best metrics as tested • It takes both pixel-based metrics (RMSE) and block-based metrics (BMR) into consideration. • Both smooth (3x3) and de-block (chou’s) show improvement for low bit-rate. EE368B