A Vision Enhancement Network for Image Quality Assessment

Abstract

With the development and update of electronic equipment, image quality assessment has become one of the hot topics. Recently, digital image processing and convolutional neural networks (CNN) have made significant progress. However, the models based on human vision characteristics and neural feedback have poor performance in previous studies. Inspired by this, we propose a CNN-based network, vision enhancement network (VE-Net). It can filter images adaptively according to the key regions. Key regions are extracted with the incentive support method from deep information learned by CNN. The adaptive filter uses Laplacian filter and Gaussian filter. Laplacian filter adopts a linear lifting algorithm, aiming to attach the image texture to the original image. Squared earth mover’s distance (EMD) loss is selected to predict the image aesthetic score distribution. VE-Net is evaluated on AVA dataset for the regression task and the classification task. Experiments show the superiority of VE-Net.