Image Denoising Network Combining Sub-Image Decomposition and Crack Repair

The image enhancement effect that is difficult to achieve by traditional methods can be easily realized by convolutional neural networks (CNN). However, the computational complexity of CNN is relatively high, making it difficult to be deployed on mobile terminals and embedded devices. Therefore, in order to reduce the computational complexity, the salt-and-pepper denoising network structure that combines sub-image decomposition with depth-separable convolution is proposed for the first time in NetA. In addition, a crack repair network is proposed for the first time to repair the cracks generated during the concatenation of sub-images in NetB. Furthermore, a construction method of image training dataset based on network self-generation has been proposed for constructing datasets of crack repair networks. The experimental results show that the computational complexity of this denoising network is 215.26 MFLOPS, which is $\mathbf{1 0}$ times smaller than that of the network built by conventional convolution. Meanwhile, acceptable denoising performance is achieved. The research in this paper will contribute to the deployment of artificial neural networks in portable mobile terminals.