Pseudo 3D-Attention for Real Image Denoising
Xin Xia, Hao Wu, Guowu Yuan · 2023
Attention mechanism has shown great potential in image noise reduction tasks. We can choose different attention mechanisms according to different noise reduction tasks to enhance the ability of network to extract corresponding information. The problem of blurring and smoothing of the denoised image in the real image denoising task is considered. This paper recommends an attention depth fusion mechanism to solve this problem. It is different from the previous use of serial or parallel channels and spatial attention mechanisms to output features directly. After parallel connection, we use the shuffle operation to achieve cross channel communication and accelerate feature fusion of the two attention mechanisms. Then use a simple MLP module to perform cross direction channel attention calculation on the acquired spatial and channel attention features. And completing the deep fusion and feature interaction of spatial attention mechanisms and channel attention mechanisms. Final we use the Cross-attention to further enhance ability of the model to extract global feature and forward long semantic information.We name this module Pseudo 3D Attention. Finally, we conduct evaluations on real image denoising benchmarks including SIDD, DND, CC15, PolyU. We proposed method achieve competitive results. In particular, PSNR of 39.71dB and SSIM of 0.961 in SIDD is achieved without extra an train set.