Toward Optical Military Targets: Benchmark Dataset and Lightweight Detection Model

With the support of artificial intelligence (AI) technology, modern warfare has developed towards unmanned and intelligent direction. Unmanned aerial vehicles (UAVs) equipped with optical sensors play a key role in battlefield situation awareness with its flexibility and cost-effectiveness. Aiming at the lack and sensitivity of military target data, this paper studies military targets from the perspective of aviation and constructs a military target dataset for target detection. The benchmark encompasses five types of targets: tank, battleship, fighter, chopper, and soldier. We also make statistical analysis on target attributes and characteristics. In addition, an improved feature enhancement module is developed and integrated into YOLOv7 baseline to realize adaptive selection of channel information. Then, the channel pruning is carried out on the improved detection model to achieve the goal of model lightweight. Evaluation on the proposed dataset shows that the complexity of the final model is reduced while maintaining the detection performance. This work provides a feasible benchmark and solution for military reconnaissance based on UAVs.