Stereoscopic image quality assessment via convolutional neural networks

Abstract

This paper mainly introduces an image quality assessment for stereoscopic images via Convolutional Neural Networks (CNN). Firstly, the left and the right view images of a stereoscopic image need to be fused in the way of Principal Component Analysis (PCA). Secondly, method of Mean Subtraction and Contrast Normalization (MSCN) is applied in the fusion images. Finally, taking non-overlapping small patches of each image as the input, the CNN can train an evaluation model between image features and Different Mean Opinion Scores (DMOS). The model can predict quality scores of image patches and we average the patch scores of each image to get the final predicted quality scores of large images.