Deep Differential Convolutional Network for Single Image Super-Resolution

The deep convolutional neural networks and residual networks have shown great success and high-quality reconstruction for single image super-resolution. It is clearly seen that among the best-known super-resolution models, deep learning-based methods demonstrate state-of-the-art performance. In this paper, we propose a deep differential convolutional network (DCN) for single image super-resolution (SRDCN). The proposed DCN is a novel convolutional network, which is composed of convolutional layers, parametric rectified linear units (PReLU), and the identity skip connection. Different from other deep learning-based methods which complete the reconstruction by learning the mapping function between low-resolution and high-resolution images, the proposed algorithm makes changes to the way of reconstruction. In the proposed network, we use DCN to obtain the reconstructed images and the differences between the low-resolution and reconstructed images in the reconstruction process. Then the differences combined with the original low-resolution image and the reconstructed image that from the last DCN are used for final reconstruction. In addition, the loss function is more rationally designed and optimized in this paper. The proposed loss function contains three parts of loss: feature loss, style loss, and mean squared error (MSE) loss. These losses will be used to supervise the structure and content of the reconstructed image. The experimental results prove that the proposed model is superior to many state-of-the-art super-resolution methods in terms of both peak signal-to-noise ratio (PSNR) and structural similarity index metrics (SSIM).