Dual Attention XYDeblur: Attention-Guided Divide-and-Conquer Strategy for Image Deblurring

Image deblurring is a fundamental yet challenging task in computer vision, aiming to restore sharp details from blurred inputs. Recent deep neural network (DNN)-based approaches have shown significant advancements, but key limitations remain. Existing methods struggle to effectively capture blur features in regions with overlapping horizontal and vertical components, leading to suboptimal restoration of fine details such as textures and edges. Moreover, conventional attention-in-attention modules often rely on sequential gating, which can attenuate fine-scale yet informative cues and reduce adaptability under spatially varying blur. To address these issues, this paper introduces the dual attention XYDeblur, a novel framework that integrates a dual-encoder-decoder architecture with complementary feature learning between the two directional streams. Unlike prior XYDeblur designs that apply directional decomposition only at the decoder stage, the proposed method performs horizontal-vertical decomposition in both the encoder and decoder, progressively refining direction-aware representations for more accurate structure reconstruction. Specifically, dual encoders separately extract blur features from horizontal and vertical components, while dual decoders branches share convolutional weights while incorporating a 90°-rotated stream, enabling efficient directional filtering without learning separate oriented kernels. In addition, the proposed parallel dual-attention module mitigates information loss induced by stacked attention by preserving complementary channel and spatial cues, thereby producing richer representations for deblurring. Extensive experiments conducted on public deblurring datasets validate the effectiveness of dual attention XYDeblur, demonstrating its ability to consistently restore sharper images and achieve superior performance over state-of-the-art methods in terms of accuracy, robustness, and computational efficiency. When evaluated on the GoPro, HIDE 1, and HIDE 2 datasets, the proposed method improves PSNR by 1.72 dB, 1.67 dB, and 2.03 dB over XYDeblur, respectively.