Video quality evaluation method based on 3D wavelet transform

摘要

A video quality evaluation method based on 3D wavelet transform utilizes 3D wavelet transform in the video quality evaluation, for transforming the group of pictures (GOP for short) of the video. By splitting the video sequence on a time axis, time-domain information of the GOPs is described, which to a certain extent solves a problem that the video time-domain information is difficult to be described, and effectively improves accuracy of objective video quality evaluation, so as to effectively improve relativity between the objective quality evaluation result and the subjective quality judged by the human eyes. For time-domain relativity between the GOPs, the method weighs the quality of the GOPs according to the motion intensity and the brightness, in such a manner that the method is able to better meet human visual characteristics.

主权项

1. A video quality evaluation method based on 3D wavelet transform, comprising steps of:
a) marking an original undistorted reference video sequence as Vref, marking a distorted video sequence as Vdis, wherein the Vref and the Vdis both comprise Nfr frames of images, wherein Nfr≧2n, n is a positive integer, and nε[3,5]; b) regarding 2n frames of images as a group of picture (GOP for short), respectively dividing the Vref and the Vdis into nGoF GOPs, marking a No. i GOP in the Vref as Grefi, marking a No. i GOP in the Vdis as Gdisi, whereinnGoF=⌊Nfr2n⌋,the symbol └ ┘ means down-rounding, and 1≦i≦nGoF;
c) applying 2-level 3D wavelet transform on each of the GOPs of the Vref, for obtaining 15 sub-band sequences corresponding to each of the GOPs, wherein the 15 sub-band sequences comprise 7 level-1 sub-band sequences and 8 level-2 sub-band sequences, each of the level-1 sub-band sequences comprises2n2frames of images, and each of the level-2 sub-band sequences comprises2n2×2frames of images;
similarly, applying the 2-level 3D wavelet transform on each of the GOPs of the Vdis, for obtaining 15 sub-band sequences corresponding to each of the GOPs, wherein the 15 sub-band sequences are 7 level-1 sub-band sequences and 8 level-2 sub-band sequences, each of the level-1 sub-band sequences comprises2n2frames of images, and each of the level-2 sub-band sequences comprises2n2×2frames of images;
d) calculating quality of each of the sub-band sequences corresponding to the GOPs of the Vdis, marking the quality of a No. j sub-band sequence corresponding to the Gdisi as Qi,j, whereinQi,j=∑k=1KSSIM(VIrefi,j,k,VIdisi,j,k)K,1≤j≤15,1≤k≤K,K represents a frame quantity of a No. j sub-band sequence corresponding to the Grefi and the No. j sub-band sequence corresponding to the Gdisi; if the No. j sub-band sequence corresponding to the Grefi and the No. j sub-band sequence corresponding to the Gdisi are both the level-1 sub-band sequences, thenK=2n2;if the No. j sub-band sequence corresponding to the Grefi and the No. j sub-band sequence corresponding to the Gdisi are both the level-2 sub-band sequences, thenK=2n2×2;VIrefi,j,k represents a No. k frame of image of the No. j sub-band sequence corresponding to the Grefi, VIdisi,j,k represents a No. k frame of image of the No. j sub-band sequence corresponding to the Gdisi, SSIM ( ) is a structural similarity function, andSSIM(VIrefi,j,k,VIdisi,j,k)=(2μrefμdis+c1)(2σref-dis+c2)(μref2+μdis2+c1)(σref2+σdis2+c2),μref represents an average value of the VIrefi,j,k, μdis represents an average value of the VIdisi,j,k, σref represents a standard deviation of the VIrefi,j,k, σdis represents a standard deviation of the VIdisi,j,k, σref-dis represents covariance between the VIrefi,j,k and the VIdisi,j,k, c1 and c2 are constants, and c1≠0, c2≠0;
e) selecting 2 sequences from the 7 level-1 sub-band sequences of each of the GOPs of the Vdis, then calculating quality of the level-1 sub-band sequences corresponding to the GOPs of the Vdis according to quality of the selected 2 sequences of the level-1 sub-band sequences corresponding to the GOPs of the Vdis, wherein for the 7 level-1 sub-band sequences corresponding to the Gdisi, supposing that a No. p1 sequence and a No. q1 sequence of the level-1 sub-band sequences are selected, then quality of the level-1 sub-band sequences corresponding to the Gdisi is marked as QLv1i, wherein QLv1i=wLv1×Qi,p1+(1−wLv1)×Qi,q1, 9≦p1≦15, 9≦q1≦15, wLv1 is a weight value of Qi,p1, the Qi,p1 represents the quality of the No. p1 sequence of the level-1 sub-band sequences corresponding to the Gdisi, Qi,q1 represents the quality of the No. q1 sequence of the level-1 sub-band sequences corresponding to the Gdisi; and selecting 2 sequences from the 8 level-2 sub-band sequences of each of the GOPs of the Vdis, then calculating quality of the level-2 sub-band sequences corresponding to the GOPs of the Vdis according to quality of the selected 2 sequences of the level-2 sub-band sequences corresponding to the GOPs of the Vdis, wherein for the 8 level-2 sub-band sequences corresponding to the Gdis supposing that a No. p2 sequence and a No. q2 sequence of the level-2 sub-band sequences are selected, then quality of the level-2 sub-band sequences corresponding to the Gdisi is marked as QLv2i, wherein QLv2i=wLv2×Qi,p2+(1−wLv2)×Qi,q2, 1≦p2≦8, 1≦q2≦8, wLv2 is a weight value of Qi,p2, the Qi,p2 represents the quality of the No. p2 sequence of the level-2 sub-band sequences corresponding to the Gdisi, Qi,q2 represents the quality of the No. q2 sequence of the level-2 sub-band sequences corresponding to the Gdisi; f) calculating quality of the GOPs of the Vdis according to the quality of the level-1 and level-2 sub-band sequences corresponding to the GOPs of the Vdis, marking the quality of the Gdisi as QLvi, wherein QLvi=wLv×QLv1i+(1−wLv)×QLv2i, wLv is a weight value of the QLv1i; and g) calculating objective evaluated quality of the Vdis according to the quality of the GOPs of the Vdis, marking the objective evaluated quality as Q, whereinQ=∑i=1nGoFwi×QLvi∑i=1nGoFwi,wi is a weight value of the QLvi.