No-reference Quality Assessment Based on Convolutional Neural Network

Abstract

How to extract image features highly correlated with visual perception is still a challenging task in No-reference image quality assessment. We aim to test the feasibility of introducing deep learning into quality assessment algorithm. In this paper, we developed a novel convolution neural network IQF-CNN that is able to learn more discriminative image quality features, and applied the learned features to predict image quality. We also used the local luminance coefficients normalization and dropout technology to improve the IQF-CNN learning ability. The proposed method can accurately measure the five common image distortions on standard benchmarks, and also improvements over the previous state of the art NR-IQA works have been demonstrated.