Convolutional neural network with gradient information for image super-resolution

Image super-resolution (SR) can help people obtain a high-resolution (HR) image through a series of low-resolution (LR) images. Deep convolutional neural network (CNN) is one of deep learning methods, and has been applied in the field of machine vision successfully. Recently, people developed a CNN model, which learns an end-to-end mapping between LR/HR images directly. However, this method only considers the local gray information of images in the training step, and we find that the experimental results produce local edge blurring. In this paper, we consider a combination of local gray information and local gradient information to reduce the local edge blurring. We calculate the 1st (first) and 2nd (second) order for each pixel to represent image edge information by gradient algorithm. Then, we combine our training samples and the gradient information together to train the SRCNN model. Finally, the experimental result shows that our method is effective by comparing with the state-of-the-art SR methods.