Enhancing defocus blur detection through dual recurrent complementary residual refinement

Longrui Li, Liqing Huang, Tianqiang Huang, Haifeng Luo

Abstract: In recent years, despite significant progress in defocus blur detection (DBD), challenges remain, particularly in distinguishing homogeneous regions and capturing fine boundary details. Most deep learning-based methods focus on capturing multilevel features through deeper or wider networks, overlooking the significance of channel and spatial information within single-level features. To address this, we propose DRCR\(^2\)Net, a Dual Recurrent Complementary Residual Refining Network comprising a channel attention branch and a spatial attention branch. These branches individually learn and capture channel and spatial information from single-level features, respectively, recurrently refining these features at each level. Additionally, a circular mutual feedback mechanism facilitates cross-enhanced residual learning and refinement between the branches. This approach directly refines channel and spatial information for single-level features, improving the detection of homogeneous regions and boundary details. Moreover, based on the complementary nature of channel and spatial information, the mutual feedback mechanism further strengthens feature expression, enabling the network to handle complex scenarios more effectively. Fusion of the branches’ outputs generates the final result. Comprehensive experiments on the CUHK and DUT datasets demonstrate that our method outperforms state-of-the-art methods, improving F-Measure by 0.013 and 0.05, S-Measure by 0.03 and 0.05, and reducing MAE by 0.014 and 0.029, respectively. Our codes are availiable at https://github.com/LLR111/DRCR2Net.

External IDs:dblp:journals/vc/LiHHL25