Infrared Object Detection Based YOLOv8s from UAV Images

Driven by artificial intelligence technology, drones have demonstrated strong perceptual capabilities. Compared to conventional images, drone-captured images offer a wide aerial perspective and exhibit strong resistance to interference due to their unique imaging mechanisms. Infrared object detection algorithm from UAV-view images can be widely applied in both military and civilian fields. In this paper, we propose a novel algorithm based on the widely-used YOLOv8s object detection algorithm model. The algorithm incorporates a lightweight Shuffle Attention (SA) module for YOLOv8s using Shuffle Units to combine spatial and channel attention mechanisms, facilitating communication across sub-features. Moreover, we also improve the original loss function by bounding box similarity comparison metric MPDIoU based on minimum point distance. While encompassing all the crucial elements considered in traditional loss functions, this metric simplifies the calculation process. Compared to YOLOv8s, the mean average precision(mAP) values achieved by our method on datasets, including DroneVehicle dataset, infrared security dataset, maritime ship dataset, and infrared aerial photography, represent increases of 1%, 0.7%, 0.4%, and 1% respectively.