Self-supervised denoising has attracted widespread attention due to its ability to train without clean images. However, noise in real-world scenarios is often spatially correlated, which causes many self-supervised algorithms based on the pixel-wise independent noise assumption to perform poorly on real-world images. Recently, asymmetric pixel-shuffle downsampling (AP) has been proposed to disrupt the spatial correlation of noise. However, downsampling introduces aliasing effects, and the post-processing to eliminate these effects can destroy the spatial structure and high-frequency details of the image, in addition to being time-consuming. In this paper, we systematically analyze downsampling-based methods and propose an Asymmetric Tunable Blind-Spot Network (AT-BSN) to address these issues. We design a blind-spot network with a freely tunable blind-spot size, using a large blind-spot during training to suppress local spatially correlated noise while minimizing damage to the global structure, and a small blind-spot during inference to minimize information loss. Moreover, we propose blind-spot self-ensemble and distillation of non-blind-spot network to further improve performance and reduce computational complexity. Experimental results demonstrate that our method achieves state-of-the-art results while comprehensively outperforming other self-supervised methods in terms of image texture maintaining, parameter count, computation cost, and inference time.

翻译：自监督去噪因无需干净图像即可训练而受到广泛关注。然而，现实场景中的噪声通常存在空间相关性，这导致许多基于像素独立噪声假设的自监督算法在真实图像上表现不佳。近年来，非对称像素重组下采样（AP）被提出用于破坏噪声的空间相关性。但下采样会引入混叠效应，而消除这些效应的后处理过程不仅耗时，还会破坏图像的空间结构与高频细节。本文系统分析了基于下采样的方法，并提出一种非对称可调盲点网络（AT-BSN）来解决这些问题。我们设计了盲点大小可自由调节的盲点网络：训练时采用大尺寸盲点以抑制局部空间相关噪声，同时最小化对全局结构的损伤；推理时采用小尺寸盲点以最小化信息损失。此外，我们提出盲点自集成与非盲点网络蒸馏方法，进一步提升性能并降低计算复杂度。实验结果表明，我们的方法在图像纹理保持、参数量、计算代价及推理时间等方面全面超越其他自监督方法，取得了最先进的性能。