AMSA-Net: Attention-based multi-scale feature aggregation network for single image dehazing

Shanqin Wang¹Mengjun Miao^1,2*Miao Zhang¹

• ¹Chuzhou Polytechnic, Chuzhou, China
• ²Qinghai Nationalities University, Xining, China

Problem: Deep learning has significantly advanced the single-image dehazing. However, many existing methods fail to adequately consider haze density and its spatial distribution, limiting further improvements in dehazing performance. Proposed solution: To address this issue, we propose an attention-based multi-scale feature aggregation network (AMSA-Net) for single-image dehazing. Method: AMSA-Net introduces a multi-scale hybrid attention feature aggregation module (MSHA-FAM) within an encoder-decoder framework to enhance the perception of haze density and spatial information, thereby enhancing dehazing effectiveness. MSHA-FAM comprises two key components: the scale-aware coordinate residual module (SCRM) and multi-scale feature refinement residual module (MSFRRM). SCRM integrates coordinate attention to capture haze density and spatial feature effectively, leading to substantial improvements in haze removal. MSFRRM extract semantic features via up/down sampling operations and enhances critical features using an improved pixel attention mechanism. In the overall MSHA-FAM pipeline, SCRM first learns haze density and spatial distribution features, which are subsequently refined by MSFRRM for more effective dehazing. Key results: The experimental results demonstrate that our proposed AMSA-Net outperforms existing methods in dehazing quality. Ablation studies further confirm the effectiveness of the proposed modules. Impact: In this work, we present AMSA-Net, which delivers competitive haze removal performance and provides a reliable visual foundation for downstream computer vision applications.