Research on image denoising based on deep learning algorithm

To solve the problem of image feature degradation and information distortion in complex noise environment, an adaptive denoising algorithm based on improved convolutional neural network (RIDCNN) is proposed in this paper. By constructing a hybrid architecture of "noise residual learning + multi-scale collaborative optimization", the edge details and texture features can be accurately preserved while eliminating additive white Gaussian noise. This algorithm adopts lightweight design idea, replaces the traditional full-connected layer with 16-layer convolution structure, and enhances local receptive field and feature multiplexing ability by combining void convolution and skip connection mechanism. In the experiment, SIDD and DND double reference data sets were used to compare six algorithms such as BM3D and DnCNN. The experimental results showed that the visual quality evaluation of the proposed algorithm showed that it could effectively suppress GGhost and maintain color consistency. The ablation experiment further verifies that the 3×3 convolution kernel and the three-layer depth are the optimal combination of parameters, and the results show that the network performance is nonlinear correlated with the number of layers. The research results of this paper provide an efficient solution for intelligent image preprocessing in real scenes.