No Reference Image Quality Assessment based on Multi-Expert Convolutional Neural Networks

Abstract

No Reference (NR) Image Quality Assessment (IQA) algorithm is capable of measuring the quality of distorted images without referencing the original images. This property is of great importance in image processing, compression, and transmission. However, due to the diversity of the distortion types and image contents, it is difficult for the existing NR IQA algorithms to be applied and maintain the best performance for all cases. To address this problem, we develop a novel NR IQA algorithm based on multi-expert convolutional neural networks (CNNs), which consists of distortion type classification, CNN based IQA algorithms and fusion algorithm. First, we present a distortion type classifier to identify the distortion type of the input image. Then, we propose a multi-expert CNN based IQA algorithms for each type of distortion. Finally, a fusion algorithm is adopted to aggregate the classification result of distortion types and multi-expert CNN based image quality predictions. The proposed algorithm has been tested on commonly used LIVE II database and a cross-dataset evaluation was carried on CSIQ database. The experimental results show that the proposed algorithm provides effective improvements for NR IQA.