LCE-YOLO: A Lightweight Detection Model for Aerial Images

• Published in: International Conference on Automation, Control and Robotics Engineering (Online) pp. 177 - 181
• Main Authors: Zhang, Zhicong, Pan, Feng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.07.2025
• Subjects: Accuracy; Adaptation models; Autonomous aerial vehicles; Computer architecture; Embedded device; Feature extraction; Lightweight; Model compression; Neck; Parameter reduction; Performance evaluation; Semantics; Visualization; YOLO
• ISSN: 2997-6278
• DOI: 10.1109/CACRE66141.2025.11119593

There is an increasing demand for object detection tasks on embedded or mobile devices via YOLO. However, due to the characteristics of its network structure, it requires high hardware configuration, and at the same time needs to ensure the detection effect, which poses challenges to the deployment of the model. To solve the problems in the above studies, a lightweight model (LCE-YOLO) is designed and proposed in this study. First, LAE is used at the neck instead of ordinary convolution to cut down parameter size while maintaining semantic information. Secondly, inspired by the cross-scale feature fusion of CNN, Through a multi-scale feature fusion mechanism, the adaptability of the model to the change of the target size is enhanced, which improves the detection performance of small-scale targets. Finally, the global information is encoded using the EMA attention mechanism module, which further aggregates pixel-level features through dimension interaction. Experimental results show that LCE-YOLO outperforms YOLOv11n, with mAP increased by 6.8%, the number of parameters reduced by 45.17%, and GFLOPS reduced by 20%.