关键词: 小波变换, 傅里叶变换, 运动图像去模糊, 生成对抗网络

Abstract: Relative motion between camera and subject causes motion image blur, losing structural and edge details. However, existing motion deblurring methods exhibit inadequate frequency domain processing, underutilization of spectral features, and neglect of high-frequency information during training. We propose a multi-domain information fusion generative adversarial network (MIFGAN) to address this issue. MIFGAN integrates a Dual-Wavelet Convolution Module (DWCM) to separate structural and edge details while expanding receptive fields, enhancing local feature extraction and detail preservation. Meanwhile, a Fourier Circular Attention Module (FCAM) applies annular attention on the Fourier spectrum to enhance critical frequency components. Additionally, the framework employs wavelet detail loss and fourier high-frequency enhancement loss to optimize detail and spectral fidelity. Experimental results of MIFGAN on multiple datasets demonstrate the promising performance of the proposed method in the task of motion image deblurring.

Key words: wavelet transform, fourier transform, motion image deblurring, GAN