Perceptual metric for image quality

Most existing efforts in image and video compression have focused on developing methods that minimize mathematically tractable distortion metrics (such as MSE). While such distortion measures were found to perform well for many applications, they do not always correlate well with the perceived quality. In this paper we present a perceptual model, which would decide which code/decoder method is the best at different code rate. The main idea of this perceptual model is to firstly decompose both the original image and the recovered image using wavelet transform, and then compare them at different subband and orientation to get perceptual PSNR We use this perceptual model to compare four different codec, i.e. JPEG, PIC4x4, PIC8x8 and Said Pearlman codec for the typical images-Bank. The results are well correlated with the human visual system. Another interesting point here is that we find that this wavelet based perceptual metric does favor the wavelet-based codec Said-Pearlman codec somehow (esp. at the lower bit rate).