No-reference Distorted Image Quality Assessment Based on Deep Learning

In this paper, a new no-reference Image Quality Assessment (IQA) algorithm for distorted image using deep learning network is proposed. The algorithm is composed of three steps. First, a multi-layer convolutional neural network is designed, which consists of four convolutional layers, four pooling layers and three fully connected layers. Second the original distortion image database is used to train the network. In order to solve the problem of insufficient training samples, this paper proposes an image segmentation method to segment the image expansion data set. Third, the trained network is used to predict the distorted image quality score. And it is not necessary to manually extract image features, it makes full use of the deep convolutional network learning ability. Experimental results on four open distorted image databases show that the proposed method predicted distorted image quality scores results have high consistency with the subjective assessment results. This method overall performance is better than other classic image quality assessment method.