A Joint Detection Method for Military Targets and Their Key Parts for UAV Images

Accurate detection of military targets and their key parts is of great significance for achieving maximum damage efficiency of attacking drones. However, traditional methods either compromise on accuracy or model efficiency, resulting in poor performance in practical applications. In this paper, we proposed a deeply coupled joint detection network with higher accuracy and smaller model size, named Joint_TK Net. Joint_TK Net explicitly models the temporal, spatial, semantic relationships between the full target detection task and key part detection in a lightweight way, thereby improving accuracy. The precise localization of the key part is intricately linked to the specific type of the target, so that we designed a class-knowledge driven feature adjustment module that utilizes the predicted class and score output by the full target detector to guide the key part detector. A context refinement and integration module is built to improve feature representation ability while reducing model parameters through aggregating global context information. We also proposed a cross-task supervision loss function to model the spatial containment relationships between full targets and key parts. Experiments on both the simulated military targets dataset and real-world dataset show that the proposed method greatly outperforms the state-of-the-art methods. The effectiveness is also verified through extensive ablation studies.