Image denoising method based on double-branch hole residual convolutional neural network
Yue Su, Xinlei Jia, Yali Peng, Shigang Liu · 2022
Relying on powerful learning and representation capabilities, Convolutional Neural Networks(CNN) have better denoising performance than traditional methods in image denoising tasks. However, these methods generally have the disadvantages of excessive smoothing and large amount of parameters. In order to overcome this shortcoming, an image denoising method based on double-branch hole residual convolution neural network is proposed in this paper. This method designs a novel network structure. By designing different voids on the branches, the detailed information at different levels of the image is fully mined. By reducing the number of convolution kerns on the branches, the number of parameters in the whole network is reduced. Compared with Denoising Convolutional Neural Network(DnCNN), the parameters are reduced by 19.6%, and the residual mechanism further accelerates the training process. On the test datasets, the proposed method not only achieved excellent performance in the two quantitative evaluation indicators of peak signal-to-noise ratio and structural similarity, but also subjectively denoised images contained richer texture details.