Hierarchical Local-Global Attention in a Multi-Scale Transformer Network for Enhanced Image Denoising

Abstract

Image denoising aims to remove noise from contaminated images. With the increasing complexity of noise in real-world scenarios, current denoising methods struggle to effectively address this challenge. This paper proposes a Multi-Scale Transformer Network (MST-Net) for image denoising. First, we introduce a novel multi-scale patch embedding strategy. In this process, noisy images are divided into patches of varying scales to capture multi-scale features. Second, we propose a Hierarchical Local-Global Attention (HLGA) mechanism in MST-Net. The proposed HLGA initially produces local attention within each scale, which is then integrated with global attention to generate the final attention map. Consequently, our MSTNet can capture long-range dependencies at multiple scales, effectively reducing complex noise in the denoising process. Additionally, we introduce a cross-scale feature fusion module to enhance information integration across different scales. Extensive experiments on standard benchmarks, including Set12, BSD68, CBSD68, and Urban100 datasets, demonstrate that the proposed MST-Net achieves state-of-theart performance. Specifically, MST-Net outperforms existing methods by up to 0.17 dB PSNR improvement on Set12 and 0.15 dB on BSD68 at higher noise levels (σ=75). Moreover, on color image datasets, MST-Net shows consistent enhancements, achieving up to 0.13 dB PSNR gain on Urban100. These results highlight the effectiveness of MST-Net in handling diverse noise patterns while maintaining a balance between computational efficiency and denoising performance. The proposed approach offers a practical solution for real-world image denoising applications.