Registration-based distortion and binocular representation for blind quality assessment of multiply-distorted stereoscopic image

Abstract

Multiply-distorted stereoscopic images are common in real-world applications. The mixture of multiple distortions results in complex binocular visual behavior of multiply-distorted stereoscopic images, making it challenging for existing blind singly-distorted stereoscopic image quality assessment (IQA) methods to obtain satisfactory results on multiply-distorted stereoscopic images. Because binocular rivalry caused by different distortions in the left and right views greatly influences the final stereoscopic image quality, we propose a registration-based distortion and binocular representation for blind quality assessment of multiply-distorted stereoscopic image in this paper. First, we employ a registration-based distortion representation to characterize the distortion in the stereoscopic image. Then we represent the binocular rivalry by merging the left and right views into a cyclopean image. Considering that the color and intensity of pixels in the RGB image can better reflect the information of the distorted image, then a grayscale cyclopean image is further converted to the color binocular representation using tone mapping. Finally, a multiply-distorted stereoscopic IQA method based on a double-stream convolutional neural network is proposed. The two subnetworks are used to extract quality features from the registration-based distortion representation and color binocular representation, respectively. Experimental results demonstrate that the proposed model outperforms the state-of-theart models on the multiply-distorted stereoscopic image databases.