Bi-path network coupling for single image super-resolution

Abstract

Recent researches have shown that deep convolutional neural networks can significantly boost the performance of single-image super-resolution (SISR). In particular, residual network and densely convolutional network can improve performance remarkably. The residual network enables feature re-usage and the dense skip connections enables new features exploration, which are both favor for feature extraction. In order to alleviate the vanishing-gradient problem in very deep convolution networks. In this paper, a bi-path network coupling is presented for SISR by combining the residual network and the dense skip connections in a very deep network. More specifically, the feature maps in the proposed network are split into two paths, one path is propagated in the form of residual connections, and another is propagated by dense skip connections. In addition, we input the feature maps obtained from the two paths into the coupling layer for feature fusion. Finally, the deconvolution layers are integrated into the network to upscale the feature map for significantly accelerating the network, that the mapping is learned from the low-resolution image to the high-resolution image directly. The proposed network was evaluated on four benchmark datasets and has achieved competing or even higher peak signal-to-noise ratio (PSNR) than most of state-of-the-art methods.