融合多注意力机制的生成对抗网络图像去模糊

Multi-Attention Fusion Generative Adversarial Network for Image Deblurring

摘要: 生活中的图像模糊通常由相机抖动、物体运动等多因素造成，且模糊核未知，这使得图像去模糊成为病态逆问题。现有的深度学习方法虽能取得较好效果，但在处理真实、复杂的模糊时，往往存在细节恢复不足和结果过于平滑的问题。为此，本文提出了一种新颖的融合多注意力机制的生成对抗网络。该网络设计了融合自适应特征激励策略与增强通道注意力方法的多尺度金字塔架构，同时提出了新的多尺度空间特征增强模块增强模型特征提取能力。我们在不同基准数据集(如GoPro和HIDE)上进行了大量实验，定性与定量结果均表明，本方法在恢复锐利边缘和真实纹理方面优于传统神经网络方法。

Abstract: Image deblurring represents a challenging ill-posed inverse problem, typically caused by multi-factorial origins such as camera shake and object motion with unknown blur kernels. While current deep learning approaches achieve considerable performance, they often exhibit limitations in reconstructing fine details and tend to produce over-smoothed outputs when handling real-world, complex blur patterns. To mitigate these issues, this paper introduces a novel generative adversarial network incorporating a multi-attention mechanism. The proposed framework employs a multi-scale pyramid architecture that integrates an adaptive feature excitation strategy with an enhanced channel attention methodology, alongside a newly designed multi-scale spatial feature enhancement module to augment the model’s representational capacity.