基于注意力机制和提示学习的图像去模糊网络

Image Deblurring Network Based on Attention Mechanism and Prompt Learning

摘要:

针对现有运动去模糊算法在边缘恢复效果不佳且易产生模糊伪影的问题，提出了一种基于注意力机制和提示学习的图像去模糊网络；结合注意力机制设计了特征融合模块，利用不同层的多尺度信息，引导网络关注于图像的边缘信息，以提高图像边缘复原质量；在解码器中引入轻量级提示模块，通过捕捉图像的全局结构信息，增强对模糊区域特征的重建能力，其中采用两个注意力分支减少了网络参数量和计算量；实验结果表明，该网络在三个公开数据集上的定量评价指标均表现优异，同时参数量和计算量具有一定竞争力，能够有效恢复图像边缘细节并减少模糊伪影。

Abstract:

To address the limitations of existing motion deblurring algorithms, such as poor edge restoration and the tendency to produce blurry artifacts, an image deblurring network based on attention mechanisms and prompt learning is proposed. The network integrates a feature fusion module designed with attention mechanisms, which utilizes multi-scale information from different layers to guide the network to focus on edge details, consequently improving edge restoration quality. A lightweight prompt module is introduced into the decoder to enhance the reconstruction ability of fuzzy region features by capturing the global structure information of the image, and two attention branches are used to reduce the amount of network parameters and calculations. Experimental results demonstrate that the network achieves superior performance on three public datasets in terms of quantitative evaluation metrics, while maintaining competitive parameter counts and computational efficiency. The proposed method effectively restores edge details and reduces blurry artifacts.