Enhanced YOLOv8 for small-object detection in multiscale UAV imagery: Innovations in detection accuracy and efficiency

Abstract

Due to the ability of drone technology to cover wide areas and reach difficult-to-access places, the detection of small targets by Unmanned Aerial Vehicles (UAVs) is crucial in various applications such as urban management planning and emergency response. This paper proposes innovations to the YOLOv8 architecture, significantly enhancing its performance in small target detection from multi-scale drone perspectives. We introduce Channel Priority Attention Dynamic Snake Convolution and Dynamic Small Object Detection Head Layer (DyHead-SODL) to improve the model's ability to capture fine details and detection accuracy. Additionally, we implemented an enhanced loss function (MPDIoU) and a Deformable Attention Transformer (DAT) to optimize detection efficiency without increasing computational burden. Experimental results on the Visdrone and RSOD datasets demonstrate significant improvements, with the proposed method increasing mAP50 by 10.5 %, mAP95 by 6.9 %, reducing localization error Eloc by 1.27, and decreasing the miss rate by 2.42 on the Visdrone dataset, outperforming existing state-of-the-art detection models while maintaining low computational complexity. The proposed method has also been tested on four large public datasets: DOTA, HRSC, LEVIR, and CARPK, demonstrating the generalization capability of our model. These advances provide effective solutions for small target detection from drone perspectives and promote the development of single-stage object detection technology. These advancements provide an effective solution for small object detection from UAVs, advancing the field of one-stage object detection technology.

(source code: https://github.com/GuccIceCream/yolov8/tree/master)