A Lightweight YOLOv8-Based Small Object Detection Algorithm in UAV Scenarios

Image-based object detection is a key enabler for UAV to achieve autonomous operations. However， images captured from a drone perspective are characterized by complex backgrounds and numerous small objects. Meanwhile， the limited onboard computational resources restrict the size of the model. Together， these two factors constrain the detection performance of models. To address the shortcomings of existing algorithms in drone aerial photography scenarios， this study proposes Drone-YOLO， a lightweight small object detection algorithm improved from YOLOv8. First， to tackle the problems of excessive parameters and high computational complexity in the YOLOv8 backbone network， this study designed a lightweight feature extraction network called CSPGhostNet by integrating the Ghost module. Through cross-stage feature map generation， it compresses the model size and efficiently extracts object features. Second， aiming at the issues of feature degradation and insufficient fusion in the neck layer of the YOLOv8 algorithm， this study proposes a multi-scale feature fusion module based on a hybrid attention mechanism. By combining channel and spatial attention， it strengthens local detailed features while enhancing the overall correlation of global features， thereby improving feature expression. Finally， a shared small object detection head was designed. Through a convolutional sharing mechanism， the computational load of the model is reduced while increasing detection speed. In addition， the regression loss function was improved by introducing a bounding box loss based on the dynamic non-monotonic focusing mechanism （Wise-IoU， WIoU）， which increased the weight of losses from small prediction boxes and made the model pay more attention to small object detection. Tests were conducted on the VisDrone2019 and MVSRD datasets. Compared with the original algorithm， the proposed algorithm performed excellently in all performance metrics： mAP@50 increased by 10.9% and 9.1%， respectively， and model parameters and computational load were reduced by 45.2% and 43.9%， respectively. In the detection of small object categories， the average precision increased by more than 13%. Compared with other popular algorithms currently available， the improved strategy proposed herein achieves an effective balance between accuracy and complexity， laying a foundation for the deployment of the algorithm on the drone edge side.