Zero-Shot Image Denoising via Adaptive Multi-Candidate Feature Alignment with Non-Local Self-Similarity

Abstract

Image denoising is the process of recovering clean images from noisy observations, which is a fundamental problem in low-level computer vision. Recent zero-shot image denoising methods don’t need external training datasets, because they just optimize directly on a single noisy image. But existing approaches mostly depend on implicit network priors or local sampling strategies, which makes it hard to exploit the abundant non-local structural redundancy, leading to poor detail preservation under complex noise conditions. To tackle these limitations, we propose a lightweight zero-shot image denoising framework called Adaptive Multi-Candidate Feature Alignment (AMCFA). This framework explicitly models non-local self-similarity by using multi-candidate feature representations. Specifically, multiple structurally similar candidates are collected for each pixel to construct a robust non-local feature representation, while an AFAM dynamically adjusts feature fusion according to local texture complexity and signal statistics. Moreover, a self-supervised random sampling method generates pairs of pseudo-training data from a single noisy image, without requiring external supervision. Extensive experiments on both synthetic and real-world benchmark datasets show that AMCFA always gets competitive denoising performance, but only requires 0.16M parameters, which gives a good balance between restoration accuracy and computational efficiency.