Dual-branch residual sparse network with serial-parallel structure for image denoising

This paper proposes a novel dual-branch residual sparse network (DRSNet) with serial-parallel structure for image denoising. In contrast to with the deep convolutional neural networks (CNNs) that only utilize the hierarchical features of noisy images, the proposed DRSNet has the advantages of depth and breadth search and attention-guided feature learning to obtain more comprehensive image feature information such as structural texture information and thus improve the model's denoising performance. The proposed DRSNet consists of two different branch sub-networks, i.e., residual sparse blocks (RSBs) and attention-guided residual sparse blocks (ARSBs), which enhance the denoising ability of the model by capturing complementary image feature information. Each of the sub-networks contains five sparse blocks and is connected by down-sampling and up-sampling operations to capture multi-scale information from local details to global context. It is worth mentioning that the proposed RSBs and ARSBs, which employ hybrid dilated convolution and residual connections can not only avoid the shortcomings of limited receptive field, large number of parameters and easy overfitting of standard convolution, but also solves the problem of low computational efficiency of dilated convolution, and realizes the balance of depth and breadth of the network. Experiments demonstrate that our proposed network model achieves excellent denoising performance.