Blind image quality assessment via deep recursive convolutional network with skip connection

The performance of traditional image quality assessment (IQA) methods are not robust, due to those methods exploit shallow hand-designed features. It has been demonstrated that deep neural network can learn more effective features compared with the traditional methods. In this paper we propose a multi-scale recursive deep neural network to accurately predict image quality. In order to learn more effective feature representations for IQA, many deep learning based works focus on using more layers and deeper network structure. However, deeper network layers introduce large numbers of parameters, which causes huge difficulty in training. The proposed recursive convolution layer ensures both the depth of the network and the light of parameters, which guarantees the convergence of training procedure. Moreover, extracting multi-scale features is the most prevalent approach in IQA. Based on this criteria, we using skip connection to combine information among layers, and it further enriches the coarse and fine features for quality assessment. The experimental results on the LIVE, CISQ and TID2013 databases show that the proposed algorithm outperforms all of the state-of-the-art methods, which verifies the effectiveness of our network architecture.