MEL-YOLO: A Novel YOLO Network With Multi-Scale, Effective, and Lightweight Methods for Small Object Detection in Aerial Images

Abstract

In recent years, deep learning has been extensively applied to small object detection, achieving significant advancements. Nevertheless, there remains substantial potential to improve both the effectiveness and efficiency of small object detection in Uncrewed aerial vehicle (UAV) images. In this article, we propose a novel YOLO model with multi-scale, effective, and lightweight methods for traffic small object detection, termed MEL-YOLO. Initially, the improved model reconstructed the multi-scale network structure, the high-resolution feature enhancement network increases size from 40× 40 to 160× 160. Secondly, the improved model incorporates a decoupled detection head, which separates the classification task from the localization task, thereby improving detection accuracy while increasing detection complexity. Furthermore, we explore the Soft-NMS algorithm to effectively mitigate small object occlusion and reduce missed detection. Finally, this study emphasizes the design of a lightweight model by introducing a model pruning strategy along with the Ghost module. We have selected four C3 modules, each with a substantial number of output channels, to integrate with the Ghost module. This method results in fewer parameters, lower GFLOPs, and higher FPS. Experiment on the VisDrone2019 public dataset, compared to the baseline YOLOv5s, our model demonstrates an improvement of 8.5% in mAP@0.5 and 6.2% increase in mAP@0.5:0.95. Furthermore, there is a 55.7% reduction in parameters, a decrease of 7.0% GFLOPs, and an increase of 11.3% in FPS. To demonstrates the versatility and robustness of MEL-YOLO model, we conducted validation experiments on the SODA10M and CCTSDB datasets, where the mAP@0.5 improved by 1.9% and 1.2%, the FPS increased by 18 frame/s and 6 frame/s, respectively. These results align with our expectations, indicating that our model is a more effective and lightweight algorithm for dense and small objects detection in aerial images.