MFEL-YOLO for small object detection in UAV aerial images

Ting Hou, Chengcai Leng, Jiaxin Wang, Zhao Pei, Jinye Peng, Irene Cheng, Anup Basu

Abstract: Object detection using Unmanned Aerial Vehicles (UAV) captured aerial images has become a research focus in recent years. However, since UAV aerial images have high resolution, large target scale variation, with most targets being small objects, it is challenging to accurately classify targets quickly and effectively. To tackle these problems, a lightweight small object detection model is proposed in this article, which is called multi-stage feature enhancement lightweight-YOLO (MFEL-YOLO). The MFEL-YOLO is optimized in three stages: (1) In the feature extraction stage, we introduce a hybrid C2F (HE-C2F) module to enhance the model feature extraction capability. To reduce the loss of feature information of small targets, a multi-branch feature fusion model (MFFM) is proposed at the end of the feature extraction stage. (2) In the cross-layer feature fusion stage, we propose an efficient path aggregation network (E-PAN), which employs a prior-based adaptive fusion strategy to replace traditional feature fusion. (3) Finally, in the object detection stage, we present a context-enhanced head (CE-Head) that employs multi-scale convolutions to capture contextual information, improving detection and classification accuracy. To validate the effectiveness of the proposed method, we conduct extensive experiments on the VisDrone 2019 and AITOD datasets. Compared with most other advanced detectors, our MFEL-YOLO achieves superior detection performance with a smaller parameter scale. The code will be available at https://github.com/kyxh1095/MFEL-YOLO-main.

External IDs:dblp:journals/eswa/HouLWPPCB25