No-Reference Stereoscopic Image Quality Assessment Based on Image Distortion and Stereo Perceptual Information

Abstract

In this paper, we propose a novel no-reference three-dimensional (3-D) image quality assessment (NR-SIQA) method, which takes three types of visual perceptual characteristics into consideration, including image distortion, depth perception, and binocular combination perception. The estimation of image distortion is based on the observation that blurriness, noisiness, and blockiness are considered as three important factors that affect image quality. Moreover, the value difference between any two neighboring pixels for any image satisfies a generalized Laplace distribution, which will be changed with various types of distortion existing in the image. Since the major difference between 2-D images and 3-D images is the depth, a disparity search algorithm is a key element for 3-D perception assessment. To address this, a robust disparity search algorithm is first designed based on the Gaussian average SSIM, followed by the generation of cyclopean image, rivalry map, depth map, and weight map. These four maps have taken the visual perceptual characteristics into consideration. Three types of features relating to image distortion, depth perception, and binocular disparity are extracted from these maps and stereo-pairs. In the end, machine learning is utilized to map these features to human perceptual score. Experimental results show that the proposed algorithm outperforms the recent full reference SIQA and NR-SIQA metrics on two publicly available stereoscopic image quality assessment databases.