基于局部特征增强的多尺度注意力机制的图像去模糊

Locally Enhanced Attention with Multi-Stage Architecture for Image Deblurring

摘要: 图像去模糊是计算机领域不可忽略的任务，其目的是为了从模糊图像中恢复更清晰的图像，从而保证在或许下游任务进行更高效的效果。基于深度学习的传统方法在图像信息恢复和边缘信息保留上效果有待提升，最近诸多研究将注意力机制融合进图像去模糊领域提升效果，但注意力有效特征提取和特征融合方面的效率有待提升，本文设计了局部增强注意力并结合多尺度架构进行图像去模糊，同时本文考虑对 rgba 多通道图像进行模型训练，保证边缘信息的有效提取。与传统深度学习方法相比提升了图像去模糊的效果。

Abstract: Image deblurring is a significant task in the computer vision field, aiming to recover sharper images from blurry inputs, thereby enhancing the effectiveness of downstream tasks. Traditional deep learning methods exhibit limitations in restoring image details and preserving edge information. Recent research has incorporated attention mecha- nisms to improve performance in image deblurring; however, the efficiency of attention mechanisms in extracting effective features and fusing them remains suboptimal. To address these challenges, this paper proposes a novel Locally Enhanced Attention mechanism integrated within a multi-stage architecture for image deblurring. Fur- thermore, the model is specifically designed to train on RGBA multi-channel images, ensuring the effective extraction of edge details. Experimental results demonstrate that the proposed approach significantly improves deblurring performance compared to traditional deep learning methods.