Masked and Shuffled Blind Spot Denoising for Real-World Images

TL;DR

This work tackles real-world image denoising under correlated noise by extending Blind Spot Denoising (BSD) into MASH, a self-supervised, single-image framework. It analyzes how masking level interacts with noise correlation and introduces Local Pixel Shuffling to decorrelate noise within small regions, coupled with an automated masking strategy that estimates noise correlation via a masking-based noise level gap. The key contributions are: 1) a detailed study of masking ratio effects under different noise correlations, 2) the adaptive masking mechanism, 3) the Local Pixel Shuffling technique, and 4) an experimental validation showing MASH achieves state-of-the-art performance among single-image/self-supervised methods on SIDD, FMDD, and PolyU datasets. The results demonstrate practical denoising improvements for real-world images and offer a pathway to robust, dataset-free denoising that adapts to the level of noise correlation.

Abstract

We introduce a novel approach to single image denoising based on the Blind Spot Denoising principle, which we call MAsked and SHuffled Blind Spot Denoising (MASH). We focus on the case of correlated noise, which often plagues real images. MASH is the result of a careful analysis to determine the relationships between the level of blindness (masking) of the input and the (unknown) noise correlation. Moreover, we introduce a shuffling technique to weaken the local correlation of noise, which in turn yields an additional denoising performance improvement. We evaluate MASH via extensive experiments on real-world noisy image datasets. We demonstrate on par or better results compared to existing self-supervised denoising methods.

Figures (7)

• Figure 1: Samples of generated noisy images depending on the spatial correlation level. From left to right: noisy image with iid noise, noisy image with moderately correlated noise and noisy image with heavily correlated noise
• Figure 2: Impact of the masking ratio $\tau$ on the generalized BSD denoising performance (PSNR). On the horizontal axis we consider several masking ratios $\tau$ and for each we train a BSD model on data with different levels of correlation $\beta$. The optimal performance of the trained model shows a strong correlation between the masking ratio an the noise correlation. Low masking benefits the training on data with iid noise and high masking benefits the training on data with highly correlated noise.
• Figure 3: Impact of the local pixel shuffling on the denoising performance (PSNR) when noise is highly correlated. The shuffling of image regions that are approximately constant destroys the noise correlation. This brings a consistent benefit across all masking ratios.
• Figure 4: (a) Original noisy image (b) Mask capturing region flatness derived from pseudo-clean prediction (c) Noisy image with local pixel permutation on flat regions.
• Figure 5: Estimated noise level based on different correlated noise magnitude and masking ratios.