Dual-branch image deblurring method based on frequency-domain awareness

Significant advances have been made in recent research on deep learning-based single-image motion deblurring algorithms, with an increasing focus on the role of information from the frequency domain for image restoration. However, deblurring remains challenging in scenarios that require detailed recovery, noise suppression, and the handling of multiple types of blur. Most currently available methods to this end focus on processing features from either the spatial or the frequency domain, which makes it difficult to effectively balance these tasks. Furthermore, these approaches often indiscriminately combine all frequency-related information, such that this leads to interactions between frequencies that generate artifacts. To address these shortcomings, we propose a dual-branch method of image deblurring in this study that is based on awareness of the frequency domain. The dual-branch fusion mechanism efficiently combines information from both the spatial and frequency domains. Moreover, the branches operate independently to overcome the limitations of current approaches that cannot simultaneously analyze features from both domains. The results of experiments on the GoPro dataset demonstrated the effectiveness of our method. It achieved a peak signal-to-noise ratio of 32.75, which was a 1.5 % improvement over the baseline model, and a structural similarity index of 0.957, marking an increase of 0.42 %. This confirms its enhanced deblurring capabilities. The codes are available at https://github.com/Liu-1018/DFNet/tree/main.