CA-BSN: Mural Image Denoising Based on Cross–Attention Blind Spot Network

Abstract

Recently, Asymmetric pixel–shuffle downsampling and Blind–Spot Network (AP-BSN) has made some progress in unsupervised image denoising. However, the method tends to damage the texture and edge information of the image when using pixel-shuffle downsampling (PD) to destroy pixel-related large-scale noise. To tackle this issue, we suggest a denoising method for mural images based on Cross Attention and Blind–Spot Network (CA-BSN). First, the input image is downsampled using PD, and after passing through a masked convolution module (MCM), the features are extracted respectively; then, a cross attention network (CAN) is constructed to fuse the extracted feature; finally, a feed-forward network (FFN) is introduced to strengthen the correlation between the feature, and the denoised processed image is output. The experimental results indicate that our proposed CA-BSN algorithm achieves a PSNR growth of 0.95 dB and 0.15 dB on the SIDD and DND datasets, respectively, compared to the AP-BSN algorithm. Furthermore, our method demonstrates a SSIM growth of 0.7% and 0.2% on the SIDD and DND datasets, respectively. The experiments show that our algorithm preserves the texture and edge details of the mural images better than AP-BSN, while also ensuring the denoising effect.