Lightweight multiscale object detection network for real-time detection in aerial photography

With the rapid advancement of drone technology, drone-based target detection has become a key research focus, but existing methods struggle with limited computing resources and real-time performance. This paper presents YOLOV8nYM, a lightweight multi-scale detection network optimized for drone imagery. Built on YOLOv8n with YulaNet as the backbone, it improves feature extraction and replaces the feature pyramid with a multi-scale fusion module for better accuracy. Deep convolution in the model head reduces computational complexity. Experiments on the VisDrone2019 dataset show YOLOV8n-YM significantly reduces parameters and computations while maintaining accuracy, enabling faster inference for real-time drone applications.