Modification Of YOLOv4-Tiny-3L Architecture to Improve the Accuracy of Airport Object Detection

Abstract

Accurate object detection in remote sensing is crucial for air traffic surveillance, border security, and disaster response. However, detecting aircraft and other objects in aerial imagery remains challenging due to small object sizes, complex backgrounds, and occlusion. YOLOv4-Tiny-3L, a lightweight object detection model, suffers from limited small object detection capabilities and suboptimal feature extraction, leading to reduced performance in high-precision tasks. This study proposes an enhanced YOLOv4-Tiny-3L architecture by integrating Cross-Stage Partial (CSP) connections and an additional multi-scale YOLO head to improve feature extraction and detection accuracy, particularly for small and densely packed objects. The modifications significantly enhance multi-scale feature representation while maintaining low computational complexity. Experimental evaluations on aircraft and remote sensing datasets demonstrate substantial performance gains, with mAP(50) improving from 0.91,77 to 0.93,23 on the Aircraft Classification dataset and from 0.97,98 to 0.98,41 on the RSOD dataset. Compared to state-of-the-art models such as YOLOv4 and Faster R-CNN, the proposed approach achieves superior accuracy while preserving efficiency, making it highly suitable for real-time applications in airport security and aerial monitoring. These findings highlight the effectiveness of CSP-based architectural enhancements in lightweight object detection models, offering a more efficient and precise alternative for remote sensing applications.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint