No-reference stereoscopic image quality assessment using 3D visual saliency maps fused with three-channel convolutional neural network

In this paper, we present a depth-perceived 3D visual saliency map and propose a no-reference stereoscopic image quality assessment (NR SIQA) algorithm using 3D visual saliency maps and convolutional neural network (CNN). Firstly, the 2D salient region of stereoscopic image is generated, and the depth saliency map is calculated, and then, they are combined to compute 3D visual saliency map by linear weighted method, which can better use depth and disparity information of 3D image. Finally, 3D visual saliency map, together with distorted stereoscopic pairs, is fed into a three-channel CNN to learn human subjective perception. We call proposed depth perception and CNN-based SIQA method DPCNN. The performances of DPCNN are evaluated over the popular LIVE 3D Phase I and LIVE 3D Phase II databases, which demonstrates to be competitive with the state-of-the-art NR SIQA algorithms.