Blind visual quality assessment for image super-resolution by convolutional neural network

Image super-resolution aims to increase the resolution of images with good visual experience. Over the past decades, there have been many image super-resolution algorithms proposed for various multimedia processing applications. However, how to evaluate the visual quality of high-resolution images generated by image super-resolution methods is still challenging. In this paper, a Convolutional Neural Network is designed to predict the visual quality of image super-resolution. The proposed network consists of two convolutional layers, two pooling layers including average, min and max pooling, three fully connected layers and one regression layer. The contribution of the proposed method is twofold. The first one is that we propose a the deep convolutional neural network to extract the high-level intrinsic features more effectively than the hand-crafted features for super-resolution images, which can be used to estimate the image quality accurately. The other is that we divide the super-resolution image into small patches, to consider the local information for the visual quality assessment of super-resolution image as well as increase the number of training data for the deep neural network. Experimental results show that the proposed metric can obtain better performance than other existing ones in visual quality assessment of image super-resolution.