Algorithm Optimization And Contrastive Learning Research For Deep Image Denoising

With the development of the semiconductor and microelectronics industries,various imaging devices are becoming more and more popular,and images have become one of the main carriers of information transmission.As one of the low-level tasks of computer vision,image denoising can effectively improve the subjective visual experience of images and improve the accuracy of downstream tasks,so it is of great significance.This paper optimizes the image denoising algorithm based on deep learning,introduces additional information and tries a new learning paradigm.The research content mainly covers the following three aspects:1.In terms of algorithm optimization,a multi-scale feature fusion mechanism consisting of a multi-scale feature aggregation module and an adaptive convolution module is proposed.Among them,the former uses improved multi-head attention to fuse features extracted from different scales,while the latter controls information at all levels in the entire mechanism,and the two complement each other to make the fused features more effective.2.In terms of additional information mining,a lightweight spatial attention module and a multi-stage frequency-domain weighted loss function are proposed.By calculating the spatial attention map in the RGB color gamut to guide the restoration of local details,a variety of intermediate results with different denoising degrees are obtained.After the final output is fused,the weighted frequency domain loss is calculated by fast Fourier transform,and the reconstruction process of difficult frequency components is focused on supervision.A better image reconstruction effect is obtained.3.In terms of new learning paradigms,an image denoising algorithm based on contrastive learning is proposed.By computing the contrastive loss simultaneously in the pixel space and the high-dimensional latent space,it overcomes the lack of detailed information of latent space features.In the pixel space,the incompletely denoised image in the early stage of the network is used as a negative sample to enhance the information strength provided by the negative sample;in the latent space,the highfrequency features that are difficult to recover are extracted by means of frequency domain transformation,and the offline feature extractor Computing the contrastive loss enhances the denoising ability of the network.The above three aspects have improved the denoising ability of the existing model to varying degrees,and the experimental results on multiple public data sets have verified the effectiveness of the above algorithm.