JND Model And Its Application In Video Coding

Traditional video coding is based mainly on statistical correlation to remove the spatial and temporal redundancy, achieving better compression results as efficiently as possible. However, as the final recipient of the video, the human visual system (HVS) contains the perception redundancy as a result of the visual properties. In order to achieve the perception redundancy, researchers have done a lot of work. Nowadays, the Just-Noticeable-Distortion (JND) model has been widely accepted, which simulates the sensitivity of HVS to luminance contrast and spatial and temporal masking effects, describing the perception redundancy quantitatively.When the changes below the JND threshold, the human eye will not perceive this change, on the contrary, changes will be detected. Applying the JND model in video coding, the areas of high JND value will be allocated less bits due to low sensitivity and the areas of small JND value will be allocated more bits due to high sensitivity. In the premise of the same bits, it will achieve better human subjective perception. With the 3D film"Avatar"on the air, the world produced a 3D heat. 3D effect is synthesized by multi-view videos, which contain a large amount of data. Due to the bandwidth limitations, how to efficiently compress the multi-view videos appears essential. Traditional multi-view coding (MVC) considers the intra-frame, inter-frame and inter-view statistic redundancy. As the single view, the JND model in multi-view or stereo could be explored, which can be applied in the MVC and achieve greater coding effect.The paper proposes a Multi-view JND (MJND) model, which explores the property of HVS to depth masking effect. The proposed model not only contains the spatial and temporal JND, but also includes the JND in depth. Compared with the other JND models without considering the depth information, the MJND model can better match the visual perception redundancy. The MJND model is then used in macroblock (MB) quantization adjustment and rate-distortion optimization in MVC. Compared with the standard MVC scheme without JND, the proposed model can get better visual quality in the case of the same bit rate.For application in the specific stereo, this paper proposes a new binocular JND (NBJND) model for the stereoscopic video which is synthesized by the left and right videos. Through the careful observation, the stereoscopic picture is divided into three parts, which are margin region, stereoscopic region and the overlapping region. The margin region and stereoscopic region are unique, these regions are used the traditional JND model in the NBJND. However, in the overlapping region, not only the spatial and temporal masking effects and luminance contrast in both views are considered, but also the interview masking from reference view into the auxiliary view is further exploited, considering the asychronism of the reference and auxiliary view, which is described as the IJND. Compared with the other JND models, the NBJND model can match the human perception redundancy of stereoscopic video better.Finally, the NBJND model is modified to be apt for the stereoscopic image. The new model combined with SSIM is then applied in the quality assessment of stereoscopic image. Experimental results show that the proposed objective quality evaluation criteria match the subjective perceived quality of stereoscopic image well.