基于Mamba与频域融合的图像去模糊网络

Image Deblurring Network Based on Mamba and Frequency Domain Fusion

摘要:

针对相机抖动或物体运动引起的图像模糊问题，提出了一种Mamba与频域融合的网络MFNet。该网络采用翻转解码器架构，将视觉Transformer的非因果建模能力与Mamba模型框架结合，通过融合频域信息提升图像去模糊性能。设计了一种非因果像素交互方法，利用注意力状态空间方程有效建模未扫描序列中语义相似像素，并通过傅里叶变换模块缓解长距离信息衰减问题。实验结果表明，MFNet在GoPro数据集上的性能超越现有主流方法，PSNR为33.43 dB，FLOPs为66.7 G，恢复精度更高且计算开销更低，能够有效去除图像模糊并恢复细节。

Abstract:

The image blurring problem caused by camera shake or object motion is addressed by proposing a Mamba and frequency-domain fusion network, MFNet. The network adopts a flip-decoder architecture, combines the non-causal modeling capabilities of the Vision Transformer with the Mamba modeling framework, and improves image deblurring performance by fusing frequency domain information. A non-causal pixel interaction method is designed to effectively model semantically similar pixels in unscanned sequences using the attention state space equation, and the issue of long-distance information decay is alleviated by the Fourier transform module. Experimental results show that MFNet outperforms existing mainstream methods on the GoPro dataset, achieving a PSNR of 33.43 dB, FLOPs of 66.7 G, higher recovery accuracy, and lower computational overhead, while effectively removing image blur and recovering details.