No-Reference Stereoscopic Image Quality Assessment Using Natural Scene Statistics

Abstract

Stereoscopic image quality assessment is an effective way to evaluate the performance of stereoscopic video systems. However, the most of existing 3D quality assessment methods cant be consistent with the subjective results caused by unconsidered the information of 3D depth or disparity. In this paper, a blind image quality assessment method for stereoscopic images is proposed using deep learning and natural scene statistics. The proposed method is composed of two stages: firstly, the 3D distorted image is classified into symmetrical or asymmetrical distortion using the characteristics of wavelet domain and disparity information of 3D image. The Deep Belief Network (DBNs) is used to classify the wavelet domain features to distortion types. Then, the mapping relationship between the NSS features and 3D image quality is established according to the distortion type. The experiment is tested on the LIVE 3D database which includes both symmetric- and asymmetric-distorted stereoscopic 3D images. Experimental results show that the proposed objective method achieves consistent stereoscopic image quality evaluation results with subjective assessment for various types of distortion, especially show its effectiveness for assessing stereoscopic imagewithcross-distortion.