Dual-Branch Complementary Blind Network for Self-Supervised Denoising in Real-World Scenarios

Self-supervised image denoising technology has become a research hotspot due to its advantage of not requiring paired training data, but it still faces the challenge of highfrequency detail loss in real noisy scenes. In this paper, a dualbranch complementary blind network (DBC-BSN) is proposed to address the inherent defects of the blind spot network (BSN) that causes edge blur and texture loss due to the masking mechanism. First, the algorithm introduces a gradient-free complementary blind branch on the basis of the traditional blind spot network, which together with the blind spot branch constitutes a dualbranch complementary blind network. The complementary blind branch does not need to pass through the masking module and retains complete information, thereby compensating for the information loss problem caused by the blind spot branch. Secondly, a complementary blind loss function is designed based on the characteristics of the dual-branch structure to optimize the entire network. Finally, a dilated convolution residual module is designed. This module uses dilated convolution to expand the network receptive field, capture the long-range dependency and long-distance structural information in the image, and effectively retains the high-frequency components of the original image through residual connections, thereby further improving the denoising accuracy. Experimental results show that this method has significant advantages in improving image quality and restoring texture details, and can effectively deal with complex noise in real scenes.