Abstract: Existing self-supervised image denoising methods based on blind-spot networks often suffer from severe loss of image information due to the limitations in network structure. To solve this problem, an innovative self-supervised image denoising method was here proposed, which improved the traditional blind-spot network into a blind-ring network, aiming to further reduce the spatial correlation of noise. Then, to address the issue of image information loss caused by traditional masking strategies, a random recovery masking strategy was here proposed, aimed at reducing information loss while enhancing the detail information of denoising results. Finally, a dual constraint loss function is proposed to prevent over-fitting of the model while effectively preserving important information in the image. Compared with the sub-optimal self-supervised image denoising method based on blind ring networks, the proposed method improved the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and image patch perception similarity (IPPS) on the SIDD validation dataset by 0.17dB, 0.007, and 0.006, respectively, verifying its superior denoising performance.

Key words: image denoising, self-supervised learning, blind-ring network, random recovery masking, dual constraint loss function

摘要： 现有的基于盲点网络的自监督图像去噪方法，常常因为网络结构的限制，导致图像信息严重损失。为解决这一难题，首先提出一种创新性的自监督图像去噪方法，将传统的盲点网络改进为盲环网络，旨在进一步降低噪声的空间相关性；然后，针对传统掩码策略导致图像信息丢失的问题，提出一种随机恢复掩码策略，旨在减少信息损失的同时，增强去噪结果的细节信息；最后，提出一种双约束损失函数，旨在防止模型过度拟合的同时，有效保留图像重要信息。相较于次优的基于盲环网络的自监督图像去噪方法，所提方法在SIDD验证数据集上的峰值信噪比(PSNR)、结构相似性(SSIM)和图像块感知相似度(IPPS)分别提高了0.17dB、0.7个百分点和0.6个百分点，验证了所提方法具有优越的去噪性能。

关键词: 图像去噪, 自监督学习, 盲环网络, 随机恢复掩码, 双约束损失函数