LiteFlex-YOLO:A lightweight small target detection network for maritime unmanned aerial vehicles

Abstract

With frequent maritime activities, the number of overboard accidents at sea has increased, and rescue delays often lead to people being killed. Unmanned Aerial Vehicles (UAVs) have the advantages of fast localization and real-time monitoring in rescue, but the images taken by UAVs have many small targets, and the detection accuracy is insufficient; at the same time, target detection algorithms are difficult to be deployed due to the limitation of computational resources of UAVs. For this reason, this paper proposes a lightweight target detection model based on YOLOv8s improvement, LiteFlex-YOLO, which aims to improve the performance of target detection in UAVs sea rescue. Firstly, the small target sensing ability of the model is enhanced by introducing the P2 small target detection layer, secondly, replacing the C2f module with the lightweight C2fCIB module reduces the computational complexity to make the model more lightweight, furthermore, the feature extraction ability of the backbone is enhanced by using the ODConv (Omni-Dimensional Dynamic Convolution); Lastly, the attention mechanism of SimAM (Simple Attention Module) is introduced to enhance the attention of the key feature information. The final experimental results showed that, LiteFlex-YOLO achieves a [email protected] of 69.5% on the SeaDronesSee dataset, which is 18.2% improvement compared to YOLOv8s, and the model parameters are reduced to 71.2% of YOLOv8s. Moreover, compared with other SOTA algorithms, LiteFlex-YOLO performs excellently in small object detection accuracy, model lightweighting, and robustness.