JND and estimated MSE based perceptual video coding for high efficiency video coding

Abstract

In this thesis, we propose a novel PVC (perceptual video coding) scheme that controls QP (quantiza-tion parameter) values based on the relation between the MSE value which is estimated with QP values and the JND values. The PVC may achieve more efficient compression than the preceding video compression technology since the PVC understands and utilizes the characteristics of the HVS (human visual system) un-like preceding video compression technologies which encode a video in terms of the statistical properties of image signal and the signal processing theory. There are several models that describe the HVS and that can be adopted to a video compression. In this thesis, we utilize the concept of the JND (just noticeable noise). The JND is one of the good properties that describe the HVS, thus, has been explored. The JND is the mini-mum difference that the HVS can perceive. That is, if distortions are smaller than the JND values in an im-age, the HVS cannot recognize the distortions, though there exist distortions. Meanwhile, the HEVC allows QP values to vary for every CU (coding unit). Since the quantization step size is closely related to visual quality of a decoded video, we can expect that by adjusting the QP values depending on the visual sensitivity, we can get a higher-visual-quality video than an ordinary HEVC encoder with the same bitrate. Therefore, in this thesis, we propose a PVC scheme that adjusts QP values with respect to the perceptual properties. The PVC scheme we propose in this thesis controls QP values by comparing the JND values and the MSE values which are estimated from the quantization step size. To do this work, we researched the manner updating a model parameter and estimating the MAD before we estimate the MSE. Besides, we take into account the variation of QP to derive the Lagrange multiplier in R-D (rate-distortion) optimization and proposed a way to compare RD (rata-distortion) costs from different CU depths. To evaluate and validate...