MEIS-YOLO: Improving YOLOv11 for Efficient Aerial Object Detection with Lightweight Design

With the advancement of aerial technologies like drones and satellites, deep learning-driven object detection has seen considerable improvements in the processing of aerial images. Nevertheless, conventional object detection algorithms continue to encounter performance limitations, particularly when handling complex backgrounds and small objects. To tackle this problem, this paper presents MEIS-YOLO, an enhanced YOLOv11-based model designed to boost both the detection accuracy and computational efficiency in aerial image processing. The core innovation of the model lies in the introduction of a Multi-scale Edge Information Selection (MEIS) module, which selects key features highly relevant to the target detection task from multi-scale features, strengthening the representation of edge information and significantly improving detection performance under conditions of small targets and complex backgrounds. Additionally, the CBRA module, which incorporates the CSP structure, optimizes the attention mechanism, further enhancing the model's detection ability and computational efficiency. To further optimize multi-scale feature fusion, this paper introduces the Asymptotic Feature Pyramid Network (AFPN). The experimental results show that MEIS-YOLO outperforms YOLOv11 on both VisDrone-2019 and DOTA datasets, especially on small target detection and complex backgrounds, with APs increasing by 4% and 8%, respectively. At the same time, FLOPs are reduced by 8%, and the number of parameters decreases by 25%, demonstrating its substantial potential for practical applications. This study provides an efficient, accurate, and lightweight solution for UAV object detection tasks.