Mamba-powered Moiré pattern suppression: A two-stage deep learning framework for high fidelity restoration

Abstract

The moiré effect remains a persistent challenge in image processing, introducing complex artifacts that degrade visual quality, particularly during image digitization and display. Although CNNs and Transformers have shown promise in moiré removal, CNNs lack the ability to model long-range dependencies, while Transformers are computationally expensive for high-resolution images. To overcome these challenges, we propose a novel two-stage moiré pattern removal framework based on the Mamba architecture. In the first stage, a Mamba-based detection network precisely identifies moiré-affected regions, leveraging its efficient long-range dependency modeling to capture intricate pattern distributions. In the second stage, a restoration network incorporates the detected moiré information as a reference to guide the removal process, leading to more targeted and effective artifact suppression. By explicitly separating detection from restoration, our method enhances both accuracy and computational efficiency. Extensive experiments on publicly available datasets demonstrate that our approach significantly outperforms state-of-the-art methods in both quantitative and qualitative evaluations, producing clearer and more detailed image reconstructions. This innovative framework establishes a new paradigm for moiré pattern removal, offering a scalable and high-performance solution for real-world applications.