LDeblur: A Lightweight Network for Single Image Deblurring
WU Wen-bo, Lei Liu, Bin Li, Jingtao Wang, Su Na, Yun Pan · Alexandria Engineering Journal · 2025
In recent years, learning-based deblurring methods have been widely applied in the field of image deblurring, achieving significant advancements. Most methods focus on stacking a large number of basic blocks in the model, which leads to high computational costs. To this end, a lightweight network for image deblurring is proposed in this paper, which has lower computational costs and comparable recovery performance. Specifically, this paper introduces an efficient hybrid feature extraction block (HFEBlock), composed of multiple dual-domain hybrid convolution blocks (DDHCBlocks) and a local–global feedforward block (LGFBlock). The DDHCBlock, which combines spatial and frequency domain features, not only significantly enhances feature extraction capabilities but also simplifies the architecture, reducing computational costs. The LGFBlock effectively integrates local and global features across multiple scales, optimizing the model’s feature representation. Furthermore, to better leverage the blur characteristics inherent in the input images, we propose a recurrent feature supervision module (RFSM). The RFSM integrates multi-scale blurry images into the backbone network, enabling the model to focus on blurred regions. Compared with recent image deblurring methods, our model reduces computational costs by more than 35%. Meanwhile, the PSNR and SSIM values reached 33.21 dB and 0.963, respectively, on the GoPro dataset.