Open Access

ARTICLE

MFCI-YOLO: Lightweight UAV Aerial Photography Small Object Detection Method Based on Multi-Scale Feature Fusion and Contextual Information

Weiguang Wang^1,2, Jincai Li¹, Mengqi Liu¹, Mengke Liu¹, Yuan Zhang¹, Jingyan Wu^1,*, Yang Liu^3,*, Junbin Lou⁴, Yixin He⁵

1 School of Information Engineering, Henan University of Science and Technology, Luoyang, China
2 Industry Research Institute of Intelligent Systems, Longmen Laboratory, Luoyang, China
3 School of Electrical and Information Engineering, Guangdong Baiyun University, Guangzhou, China
4 College of Mechanical Engineering, Jiaxing University, Jiaxing, China
5 College of Information Science and Engineering, Jiaxing University, Jiaxing, China

* Corresponding Authors: Jingyan Wu. Email: ; Yang Liu. Email:

(This article belongs to the Special Issue: Aerial Innovation Spectrum: All-Domain Research in UAV Communication, Navigation, and Autonomy)

Computers, Materials & Continua 2026, 88(2), 82 https://doi.org/10.32604/cmc.2026.080341

Received 07 February 2026; Accepted 08 May 2026; Issue published 15 June 2026

Abstract

To improve the accuracy of small object feature detection in complex backgrounds for Unmanned Aerial Vehicle (UAV) aerial photography and reduce computational complexity, we propose the lightweight UAV aerial photography small object detection method based on multi-scale feature fusion and contextual information. Firstly, by introducing the grouped content-aware reassembly (GCA) operator and designing lightweight pinwheel context convolution (LPConv), we extend the feature fusion path to the P2 layer, constructing a lightweight multi-scale feature fusion network (SG-PANet). Through the decoupling of fine-grained small object features and background interference features by the GCA operator, combined with the anisotropic receptive field constructed by LPConv, our proposed method can effectively preserve the geometric details of small objects. Furthermore, we introduce the cross-stage dense feature refinement (CSPStage) module as the pre-refining unit of the detection head, and use the full history state awareness mechanism to strengthen feature reuse and gradient propagation to solve the problem of feature degradation across layers. We utilize the Wise-IoU v3 loss function to dynamically optimize the gradient gains of high-quality and low-quality samples, thereby enhancing the detection accuracy and convergence speed of the proposed method in complex scenarios. Finally, we verified the superiority and generalization of the proposed method on the VisDrone2019 dataset and DOTAv1.5 dataset. The results show that compared with YOLOv11n, MFCI-YOLO’s detection mAP50-95 increased by 11.1%, small object mAP50 increased by 16.1%, and mAP50 reached 80.3%. It provides a practical solution for detecting small objects in dense scenes.

---

Keywords

---