SRNet: Sparse representation-based network for image denoising

Abstract

Image denoising is a foundational but important task in image processing. Various traditional methods have been proposed to remove noise from noisy images. Sparse representation (SR) is a representative image denoising method. Although with good denoising effects, sparse representation suffers from performance bottlenecks and large time consumption. Recently, deep learning has demonstrated excellent ability in image denoising. Therefore, we consider combining the sparse representation with deep learning to make this traditional model more effective and efficient. The sparse representation-based network (SRNet) is proposed by embedding the convolutional neural network (CNN) into the sparse representation framework, in which all the parameters are learned by training. The iterative optimization process of sparse coding is unrolled into a network with several similar phases. In each phase, two subnetworks are designed with the multiscale residual block (MSR-block) to model the updating of sparse coefficient and image, respectively. The experimental results show that compared with the traditional sparse representation denoising methods, the proposed method can significantly reduce time consumption and improve denoising performance, especially in terms of reconstructing textures. The PyTorch code and models of the proposed SRNet can be released at https://github.com/JiechaoSheng/SRNet.