An Improved YOLOv8-Based Method for Small Object Detection in UAV perspective

Abstract

The target objects in UAV aerial images are usually smaller targets, and the detection of such small targets is an important research area. Although the progress of target detection is constantly improving thanks to the advancement of deep learning technology, the accuracy of small target recognition in images still poses a great challenge. This study proposes an improved YOLOv8 algorithm to improve the performance of the small target object detection algorithm. This method proposes a C2F-DCNv2 module that integrates deformable convolutional network v2 (DCNv2) to replace the C2F module of the original backbone part; in addition, a Dynamic Head (DyHead) with a self-attention mechanism is used on the head) replaced the original detection head. Through training and testing in the VisDrones2019 data set, it is shown that the method proposed in this article reached 37.9% in the mAp50 indicator in the verification data set, and the average detection speed was 33.7 FPS. Compared with the results of the baseline model, the results increased by 3.6%. Experimental results show that the target detection algorithm proposed in this article significantly improves the recognition effect of small targets in UAV aerial images.