A Novel Multi-Scale Adaptive Convolutional Network for Single Image Super-Resolution

Abstract

Recently, the methods based on deep convolutional neural networks have shown superior performance on image restoration. By using a faster and deeper network, the existing methods have achieved the breakthroughs in speed and accuracy. But, they still work hard to recover sharper edge and finer texture details for the large upscaling factor. In this paper, we propose a novel multi-scale adaptive convolutional network to solve this problem. The proposed network is mainly built on the multi-scale adaptive convolutional block which is composed of three different scales sub-blocks. Each sub-block consists of two convolutional layers, one parametric rectified linear unit layer, and some adaptive shortcuts. Different scaling parameters are assigned to these shortcuts to extract and connect diversified and complicated features adaptively. Due to the special design of MACB, the proposed algorithm can enhance the combination of various levels features and provide different ranges of image context for super-resolution (SR) reconstruction. Besides the stack of MACB, the skip connections with identity mapping are used to further aggregate the features of two specific layers in the proposed network. Moreover, in order to infer photo-realistic natural images, a perceptual loss function is proposed to supervise the reconstruction, which consists of four parts of loss: feature loss, style loss, total variation loss, and pixel loss. These losses push our reconstructed image to the target image. The experimental results on the public benchmark datasets demonstrate that the proposed algorithm achieves the best super-resolution reconstruction among the state-of-the-art SR methods.