O-Net: Bidirectional Encoding and Decoding Architecture for Infrared Small Target Detection

Infrared small target detection faces significant challenges in effectively utilizing shallow and deep features while mitigating spatial detail degradation during sampling. To address these issues, we propose O-Net, a novel network architecture featuring a bidirectional encoder-decoder pathway that effectively integrates shallow and deep information of infrared small targets. The proposed O-shape dense connection further alleviates spatial detail degradation through multi-level feature interaction and reuse. First, we introduce the adaptive pixel-wise cooperative mechanism, which enhances the encoder’s feature extraction capability by synergistically extracting and fusing local and global features. Second, we design the adaptive dual perception attention module, which dynamically focuses on key features and regions of infrared small targets by combining channel and spatial attention mechanisms, thereby improving the model’s sensitivity to target information. Finally, the O-shape dense connection strengthens the model’s contextual awareness, effectively addressing challenges such as low contrast and noise interference in complex backgrounds. Experimental results on the NUAA-SIRST and IRSTD-1K datasets demonstrate that our method outperforms state-of-the-art approaches. The code for this work is publicly available at https://github.com/liuyb77/O-Net.