Lightweight target detection based on improved YOLOv5n for unmanned aerial vehicle infrared imagery

Abstract Target detection based on unmanned aerial vehicle (UAV) in complex scenes has increasingly become a research hotspot. This paper proposes a new lightweight infrared target detection algorithm based on YOLOv5n for UAV imagery. The proposed algorithm increases the detection accuracy by improving the network structure and loss function. First, a lightweight FasterNet block is introduced into the backbone network to amplify and refine the feature information, and meanwhile the network depth is increased through block stacking to enhance the feature extraction capability. Second, a new module called efficient convolutional block (ECB) is designed and then integrated into the neck network to strengthen the feature fusion capability, which is beneficial for complex environments including dense and small targets. Finally, the complete intersection over union (CIoU) in the original loss function is replaced by the minimum point distance intersection over union (MPDIoU) to effectively estimate the positional error of the target bounding box, so as to improve the localization ability. Experimental results demonstrate that the proposed network contains only 3.1M parameters yet achieves significant performance gains over the baseline across two public benchmarks: on the DroneVehicle dataset, it improves mAP@0.5 by 4.0% and mAP@0.5:0.95 by 3.6%; similarly, on the HIT-UAV dataset, it achieves 5.0% and 1.9% improvements in mAP@0.5 and mAP@0.5:0.95 respectively. Furthermore, our method outperforms several state-of-the-art lightweight networks in consideration of detection performance and resource requirements, demonstrating the potential of the proposed algorithm for UAV visual perception tasks in infrared scenes, such as disaster rescue, traffic monitoring and wildfire detection.