Exploring Efficient Asymmetric Blind-Spots for Self-Supervised Denoising in Real-World Scenarios

TL;DR

The paper tackles real-world self-supervised denoising under spatially correlated noise by arguing that training at the original input resolution with asymmetric blind-spots is essential for texture-rich restoration. It introduces Asymmetric Tunable Blind-Spot Network (AT-BSN), whose blind-spot size can be tuned during training and inference to balance noise suppression and local spatial preservation, and further leverages a blind-spots based multi-teacher distillation strategy to distill a lightweight student from multiple teacher networks. Experimental results on SIDD and DND show state-of-the-art performance among self-supervised methods with lower computational overhead and superior texture detail preservation, validating the practical viability of the approach. The work provides a principled framework for real-world denoising without clean data, combining full-resolution training, asymmetric design, and efficient distillation to achieve strong denoising quality and efficiency.

Abstract

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 that assume pixel-wise independent noise to perform poorly. Recent works have attempted to break noise correlation with downsampling or neighborhood masking. However, denoising on downsampled subgraphs can lead to aliasing effects and loss of details due to a lower sampling rate. Furthermore, the neighborhood masking methods either come with high computational complexity or do not consider local spatial preservation during inference. Through the analysis of existing methods, we point out that the key to obtaining high-quality and texture-rich results in real-world self-supervised denoising tasks is to train at the original input resolution structure and use asymmetric operations during training and inference. Based on this, we propose Asymmetric Tunable Blind-Spot Network (AT-BSN), where the blind-spot size can be freely adjusted, thus better balancing noise correlation suppression and image local spatial destruction during training and inference. In addition, we regard the pre-trained AT-BSN as a meta-teacher network capable of generating various teacher networks by sampling different blind-spots. We propose a blind-spot based multi-teacher distillation strategy to distill a lightweight network, significantly improving performance. Experimental results on multiple datasets prove that our method achieves state-of-the-art, and is superior to other self-supervised algorithms in terms of computational overhead and visual effects.

Figures (17)

• Figure 1: Comparisons of our AT-BSN with other methods. Our method recovers more high frequency texture details.
• Figure 2: Three kinds of asymmetric operations during training and inference. Our scheme can flexibly tune the blind-spot size to meet the requirements of training and inference, achieving a balance between noise correlation suppression and local spatial destruction.
• Figure 3: Effective Receptive Field analysis of AP-BSN and our AT-BSN. Each column has the same center blind-spot size. The downsampling operation of AP-BSN can cause aliasing effects.
• Figure 4: Overview of the proposed AT-BSN framework. We employ asymmetric blind-spots for training and inference to balance the suppression of noise spatial correlation and local information preservation. We regard the trained AT-BSN as a meta-teacher network, generate multiple teacher networks by sampling different blind-spots, and execute multi-teacher distillation on a lightweight network.
• Figure 5: Implementation principle of the tunable blind-spot.