Camouflage soldier object detection network based on the attention mechanism and pyramidal feature shrinking

Due to the high level of information similarity between camouflage soldier objects and their background, traditional deep learning-based object detection networks encounter distinct error detection rates and miss detection rates when attempting to detect camouflage soldiers. To address these challenges, we proposed a camouflage soldier object detection network (AFSNet) based on attention mechanism and multi-scale feature fusion strategy. We employed an attention module to enhance the network’s capability for feature extraction. Furthermore, we proposed a novel strategy for multi-scale feature fusion based on pyramidal feature shrinking, aiming to mitigate interference caused by interpolation and prevent information loss resulting from pooling during the process of feature fusion. Moreover, we introduced a novel information handle module that enhances the network’s capability for feature fusion by regulating the information transmission pathway. Experiments demonstrated that our network exhibits a better camouflage object detection performance than state-of-arts networks. Compared to YOLOv7, our network can achieve 93 \(\%\) $\%$ AP, which is increased by 6.7 \(\%\) $\%$ with almost no computation overhead.