Infrared maritime small target detection network based on attention and partial learning convolution

Infrared maritime targets (IMTs) captured by infrared sensors lack distinct detailed features and have a low signal-to-noise ratio, which renders the detection of such targets more challenging than typical target detection tasks. To improve detection accuracy, this manuscript exploits the characteristics of IMTs and proposes a detection network based on attention and partial learning convolution (APLCnet). Firstly, a projected-global self-attention module (PGSAM) is designed based on the projected saliency of IMTs. By encoding long-range features, PGSAM effectively captures the contextual relationships between targets and their backgrounds. Secondly, an attention module based on the local-wide convolutional block attention module (LWCBAM) is devised. By introducing local attention and attention over a larger receptive field, this module helps the network highlight important spatial and channel features after feature fusion, preventing critical weak targets from being overwhelmed during the fusion process. Additionally, inspired by traditional edge detection operators, a partial learning convolutional module (PLCM) is designed and applied to shallow features, which enhances the focus on small IMTs, improving the saliency of weak and small IMTs. Experimental results demonstrate that the proposed network effectively improves the accuracy of IMT detection while maintaining higher robustness.