BinaryDemoire: Moiré-Aware Binarization for Image Demoiréing

TL;DR

This work tackles image demoiréing under extreme model compression by binarizing most network weights and activations. It introduces two novel modules, the moiré-aware binary gate (MABG) and the shuffle-grouped residual adapter (SGRA), to adapt binarized activations to the frequency-structured moiré degradations and to maintain effective residual flow during channel dimensional changes. Built on an ESDNet backbone, BinaryDemoire keeps the first and last layers full-precision while binarizing the intermediates, with MABG modulating channel-wise gates and SGRA aligning shortcuts through structured projections and interleaved mixing. Experiments on four benchmarks show BinaryDemoire achieves restoration quality comparable to or better than full-precision models while dramatically reducing parameters and FLOPs, demonstrating strong potential for deployable edge solutions.

Abstract

Image demoiréing aims to remove structured moiré artifacts in recaptured imagery, where degradations are highly frequency-dependent and vary across scales and directions. While recent deep networks achieve high-quality restoration, their full-precision designs remain costly for deployment. Binarization offers an extreme compression regime by quantizing both activations and weights to 1-bit. Yet, it has been rarely studied for demoiréing and performs poorly when naively applied. In this work, we propose BinaryDemoire, a binarized demoiréing framework that explicitly accommodates the frequency structure of moiré degradations. First, we introduce a moiré-aware binary gate (MABG) that extracts lightweight frequency descriptors together with activation statistics. It predicts channel-wise gating coefficients to condition the aggregation of binary convolution responses. Second, we design a shuffle-grouped residual adapter (SGRA) that performs structured sparse shortcut alignment. It further integrates interleaved mixing to promote information exchange across different channel partitions. Extensive experiments on four benchmarks demonstrate that the proposed BinaryDemoire surpasses current binarization methods. Code: https://github.com/zhengchen1999/BinaryDemoire.

Figures (7)

• Figure 1: Performance and efficiency comparison between binarization methods and the full-precision baseline model (ESDNet yu2022towards). The results are evaluated on the UHDM dataset.
• Figure 2: Visual comparison of binarization methods. Our proposed BinaryDemoire outperforms other methods with accurate results.
• Figure 3: Overview of the BinaryDemoire framework. The BinaryDemoire contains two modules. The first is moiré-aware binary gate (MABG), which adapts binary activation through frequency and statistical descriptors to effectively modulate the binary gate values. The second module is the shuffle-grouped residual adapter (SGRA), which provides lightweight residual alignment across channel and spatial dimensions, enabling effective feature exchange and preserving critical information during the demoiréing process.
• Figure 4: Channel-wise gate values visualization. We visualize the gate values corresponding to different channels, along with their associated binarized feature maps. For channels containing richer high-frequency details, the module learns larger gating responses to increase their contribution. This adaptive strategy highlights informative features and filters out less useful information.
• Figure 5: Compare of activation distribution with (w/, right) and without (w/o, left) the interleaved mixing operation. For ease of illustration, we only present the activation distributions of the first three groups.