A Small Object Detection Model for Improved YOLOv8 for UAV Aerial Photography Scenarios

Abstract

In order to solve the problem of poor detection performance caused by a large number of small targets, occlusion, and target blur in drone aerial image object detection, we optimized the YOLOv8 model to produce a target detection model for aerial photography scenarios. Firstly, we adopted and improved the RepLKNet architecture to enhance the backbone of YOLOv8 network, using several large kernel convolutions instead of traditional VGG network to produce a larger effective receptive field; secondly, we chose MPDIoU as the bounding box loss, further simplifying the calculation by fully exploiting the geometric characteristics of horizontal rectangles; finally, based on spatial reconstruction and channel reconstruction methods, we improved the C2F module and reduced the redundancy of space and channels in the network by optimizing the SCConv module, thus reducing the computational cost and model size. We trained and validated using the VISDrone2019 dataset. The experimental results show that the enhanced model has an average accuracy 5.4% higher than the baseline model and an mAP0.5 value 5.8% higher. Compared with other popular target detection models, our model exhibits superior performance.