Complementary Blind-Spot Network for Self-Supervised Real Image Denoising

Recently, self-supervised denoising methods have attracted significant attention due to the considerable challenge posed by constructing a large-scale real noise dataset for supervised training. The most representative self-supervised denoisers are based on blind-spot networks (BSNs), which exclude the central pixel of receptive field. However, excluding any input pixel potentially leads to the loss of vital information required for accurate predictions, especially when the excluded pixel corresponds to the output position. In addition, a standard BSN has struggled to effectively reduce real-world noise due to the spatial correlation of noise, though it makes the significant results with independently distributed synthetic noise. In this paper, we propose a novel self-supervised real-world image denoising framework called Complementary-BSN based on two reciprocal branches (Mask-Map branch and Enhanced-PD-BSN branch) with an efficient loss function to employ the pixels information ignored by masked convolution and provide additional optimization target for self-supervised output. Specifically, we exploit a block-wise random-placing (BRP) scheme for further weaken the noisy correlation to avoid the illusion of image structure recovery due to existing complex noise and make Complementary-BSN more suitable for real noise. Additionally, we develop an efficient strategy (multi-stride PD (MPD)) to fuse multiple PD strides for inference, narrowing the restoration gap between textural and flat regions. Extensive experiments on real-world datasets demonstrate that our method achieves superior performance to other state-of-the-art (SOTA) self-supervised denoising methods. The code is available at https://github.com/cuijin7382/Complementary-BSN.