<p>Defocus image deblurring aims to reconstruct sharp images from defocused inputs. Although deep learning-based methods have achieved significant progress, existing algorithms still struggle to effectively extract critical details from highly redundant feature representations and to achieve clear structural and textural reconstruction in complex texture regions. To address these issues, we propose a novel W-shaped network architecture that departs from the conventional U-shaped paradigm. This design comprises two core components: the Focused Residual Fusion Module (FRFM) and the Multi-Scale Transposed Attention (MSTA) module. By establishing a collaborative processing pipeline of “feature refinement, feature distillation, and information integration,” our model effectively reconstructs blurred image structures with complex spatial variations. To comprehensively validate the model’s performance, we conducted systematic evaluations on multiple real-world datasets. On public benchmarks including DPDD (Dual-Pixel Defocus Deblurring), RealDOF (Real Depth of Field), and RTF (Regression Tree Fields), our method outperforms current state-of-the-art approaches in quantitative metrics. Furthermore, it demonstrates strong generalization capability. On the COCO128 (Common Objects in Context 128) dataset, images processed by our method significantly enhance the performance of the YOLOv11 object detection model, effectively validating its empowering role in downstream high-level vision tasks. Additionally, on our self-constructed real-world no-reference dataset SUSE-DBD, ours achieves leading results in both no-reference image quality assessment and object detection, further highlighting its practicality and robustness in real-world scenarios. The code and the self-constructed SUSE-DBD dataset will be made publicly available upon acceptance of the paper.</p>