WaMaIR: Image Restoration via Multiscale Wavelet Convolutions and Mamba-based Channel Modeling with Texture Enhancement

TL;DR

WaMaIR addresses texture detail preservation in image restoration by integrating Global Multiscale Wavelet Transform Convolutions (GMWTConvs) for expanded receptive fields, a Mamba-Based Channel-Aware Module (MCAM) for robust cross-channel dependencies, and Multiscale Texture Enhancement Loss (MTELoss) to guide multiscale texture reconstruction. The method combines a U-shaped CNN backbone with wavelet-based multiscale feature extraction, dual-path channel attention, and a balanced loss to preserve texture across scales while remaining computationally efficient. Extensive experiments across dehazing, deraining, and desnowing demonstrate state-of-the-art performance and superior texture fidelity, supported by ablations that validate the contributions of GMWTConvs, MCAM, and MTELoss. The approach offers practical impact for real-world restoration tasks requiring high-fidelity texture recovery with scalable computation.

Abstract

Image restoration is a fundamental and challenging task in computer vision, where CNN-based frameworks demonstrate significant computational efficiency. However, previous CNN-based methods often face challenges in adequately restoring fine texture details, which are limited by the small receptive field of CNN structures and the lack of channel feature modeling. In this paper, we propose WaMaIR, which is a novel framework with a large receptive field for image perception and improves the reconstruction of texture details in restored images. Specifically, we introduce the Global Multiscale Wavelet Transform Convolutions (GMWTConvs) for expandding the receptive field to extract image features, preserving and enriching texture features in model inputs. Meanwhile, we propose the Mamba-Based Channel-Aware Module (MCAM), explicitly designed to capture long-range dependencies within feature channels, which enhancing the model sensitivity to color, edges, and texture information. Additionally, we propose Multiscale Texture Enhancement Loss (MTELoss) for image restoration to guide the model in preserving detailed texture structures effectively. Extensive experiments confirm that WaMaIR outperforms state-of-the-art methods, achieving better image restoration and efficient computational performance of the model.

Figures (5)

• Figure 1: The architecture of our WaMaIR. (a) WaMaIR adopts a U-shaped structure, consisting of several CNN Blocks and three GMWTConvs, and employs a multi-input multi-output strategy. (b) GMWTConvs first extract shallow features through four convolutional layers, then use WTConv to capture features in different frequency domains. (c) CNNBlock contains multiple residual connections, with our proposed MCAM added at the end. (d) MCAM enhances Mamba's channel feature extraction capability through AP and MP. (e) The structure of Mamba.
• Figure 2: Visualization results of degraded image after wavelet transform. In our work, we employ the Haar wavelet transform, which more effectively captures image textures after decomposition. Specifically, LL represents the overall contour of the image, LH reflects vertical edge features, HL reflects horizontal edge features, and HH represents diagonal details in the image.
• Figure 3: Image dehazing comparisions on the SOTS-indoorli2018benchmarking dataset
• Figure 4: Image deraining comparisons on the Test100 zhang2019image dataset.
• Figure 5: Image desnowing comparisons on the Snow100K liu2018desnownet dataset.