Research On DCT-based Just Noticeable Distortion Model Considering Visual Perception Characteristics

Images/videos are extensively used in various multimedia businesses and have gradually integrated into people’s daily life.In order to improve the subjective quality of multimedia,it has made great progress in image/video compression,image/video transmission and robust transmission.Because the human eye is generally the last receiver of the image/video,how to describe the perception characteristics of the human eye more accurately and effectively has aroused great interest in academia and industry.Human Visual System(HVS)is unable to perceive pixel changes below a certain visibility threshold due to its potential physiological and psychological mechanism.This threshold is called Just Noticeable Distortion(JND)threshold.The JND model reveals the lowest limitation of the perceptible distortion of the human visual system to image/video,so it is commonly explored in image/video compression and transmission.Since the JND model in DCT domain can be directly applied to the field of image/video compression based on DCT transformation,this article focuses on the JND threshold estimation model in DCT domain.The existing DCT-based JND models mainly have problems,such as low threshold accuracy and cross-domain operations.In order to solve the deficiencies in the existing DCT-based JND model,the gray image characteristics and human eye visual characteristics need to be perceived.In this paper,the perception mechanism of the human visual system is deeply studied.Combining the entropy masking effect,the visual attention mechanism and the foveated masking effect,a more accurate JND threshold estimation model in DCT domain is explored for gray images.The main research content and innovation in this article are as follows:(1)A threshold estimation model of JND in DCT domain based on the entropy masking effect is proposed.First of all,starting from the free-energy theory and Bayesian brain inference mechanism,the texture-energy similarity model of DCT blocks is constructed by fully considering the correlation between DCT blocks.And an autoregressive model based on the texture-energy similarity in DCT domain is designed to simulate the spontaneous prediction behavior of the human visual system.Then,the mapping relationship between human visual perception and prediction residual is explored to obtain block-level disorder,and the entropy masking effect is modeled as a JND threshold modulation factor on the disorder.Finally,on this basis,the JND model based on the entropy masking in DCT domain is proposed by fusing the spatial contrast sensitivity function,the luminance adaptive masking and the contrast masking.The subjective and objective experimental results show that the proposed model can eliminate more perceptual redundancy while avoiding cross-domain operations,and its performance is better than the existing JND model in DCT domain.(2)A threshold estimation model of JND in DCT domain based on the combined effect of visual attention and foveated masking is proposed.First,the top-down subjective factor and bottom-up objective factor are considered at the same time.The brightness feature,texture feature and focus feature extracted from DCT coefficients are fused to build a visual saliency model in DCT domain.The fixation points are predicted according to the visual saliency map.Secondly,we analyze the influence relationship between fixation points,and propose the competitive fixation intensity that is more in line with the human visual system.Then,the competitive fixation intensity is added to the foveated masking effect to improve the accuracy of contrast sensitivity,and a JND threshold adjustment factor based on the combined effect of visual attention and foveated masking is constructed.Eventually,by fusing the above-mentioned JND threshold estimation model based on the entropy masking effect,a JND threshold estimation model based on the combined effect of visual attention and foveated masking in DCT domain is proposed.Comparative experiments demonstrate that the proposed model further improves the DCT-based JND threshold estimation accuracy while avoiding cross-domain operations...