A Multi-scale Dynamic Fusion Network for Lightweight Small Target Detection in Unmanned Aerial Vehicles

In unmanned aerial vehicle (UAV) aerial images, the slight differences in target size and complex lighting conditions lead to insufficient detection accuracy. Moreover, the constrained computational resources of edge devices hinder real-time detection performance. The lack of small target features and the interference of complex backgrounds limit the recognition ability, seriously affecting the accuracy and efficiency of UAV inspection. However, the existing YOLO series algorithms have limitations in terms of lightweight and multi-scale adaptability. To solve this problem, this study proposes a lightweight object detection algorithm specifically designed for UAV platforms. This algorithm combines shallow convolution and deep dynamic convolution, enhancing multi-scale feature fusion. Meanwhile, by adopting the normalization and deep divisible volume product techniques, the computational complexity was effectively reduced by 73%. Furthermore, through channel recombination and the dual-path divide-and-conquer strategy, the feature expression ability of small targets has been enhanced. The experimental results show that, compared with the baseline algorithm, the improved algorithm reduces GFLOPs by nearly 27% and Parameters by 23.8% on the TinyPerson dataset. The improvement of these indicators has greatly enhanced the detection performance of the model, enabling it to perform outstandingly when carrying out target detection tasks on unmanned aerial vehicle platforms with limited computing power.