An Infrared Image Denoising Method Based on Wavelet Dual-Branch Diffusion Model

Removing noise from infrared images is of great significance in various fields. The existing infrared image denoising methods have the problem of an imbalance between noise suppression and detail preservation. In this article, we propose a wavelet dual-branch diffusion infrared image denoising method, namely WD-Diff. First, a diffusion process with noisy images as the endpoint is constructed. In the reverse recovery process, we propose a new image restoration network. Among them, the low-frequency restoration network (LFRNet) and the high-frequency restoration network (HFRNet) are designed to restore wavelet low-frequency and high-frequency degradation subgraphs, respectively. The two networks are separately trained under a two-step training strategy and a single-step training strategy. Second, the self-attention residual module (SARM) and the equidistant convolution residual module (ECRM) are specially designed and applied to the LFRNet and HFRNet, respectively. Through the self-attention mechanism and residual learning, the SARM can effectively extract low-frequency feature information. By constructing an equidistant convolution large kernel, the ECRM can accurately remove high-frequency noise. Third, we design a high-to-low-frequency fusion Transformer network (HLFTNet), which achieves high-to-low-frequency feature fusion through a spatiotemporal cross-fusion attention mechanism. Finally, extensive comparative and ablation experiments have demonstrated the superiority and effectiveness of the proposed method.