Research On Visual Perception Based Stereoscopic Image Quality Objective Assessment

With the great development of some decisive technologies such as video capturing, coding, networktransmission and naked-eye displaying, stereoscopic video has been in depth in people’s lives. The processof video transmission and video compression will inevitably bring some distortions, which will influencethe human eyes’ visual comfort seriously. Stereoscopic image quality assessment tends to be more andmore essential and necessary because of the realism of stereoscopic images gradually. This paper buildsthree effective models based on the perspective binocular vision characteristics to evaluate the imagequality. The main research contents are as follows:1) First，given that the traditional image quality assessment (IQA) mainly focus on interestedregions in regardless of visual sensitivity of human eyes, the paper proposes a new IQA based on the visualsensitivity, which is supported by just noticeable distortion (JND) model. According to the characteristicsof human eyes, the method extracts the sensitive and distorted regions. Then the original and distortedimages are evaluated by structure similarity.2) Second, in most cases, researchers only take luminance component into consideration anddisregard the color component. An IQA for stereoscopic color images is proposed based on the colorcomponent channel integration and visual threshold. Visual threshold characteristics of human eyes areused to determine whether the distortion of a stereoscopic image can be perceived. Singular values ofdifference maps between original and distorted images with respect to the left and right views are comparedwith ones of the visual threshold map of original images. Then, considering the binocular just noticeabledifference (BJND), the method assesses the quality of stereo perception. The experimental results onstereoscopic images show that the proposed objective quality is highly consistent with its subjectivequality.3) Finally, in practice, it is true that the information of original image is constantly inaccessible.In order to reduce the dependency of the original image, the paper proposes a reduced reference (RR) IQAbased on the edge strength compressive sensing. In view of the RR method consistent with the subjectiveperception in LIVE2D database, another RR IQA based on the entropic differencing of the wavelettransform coefficients is proposed. The method fuses the left and right views, and then it is verified in twopublic stereoscopic image databases (LIVE Phase1, NBU Symmetric). The experimental results show thatthe accuracy is highly competitive to the popular IQA methods.