YOLO-SWIN HYBRID MODEL FOR ENHANCED SMALL OBJECT DETECTION IN AERIAL IMAGES

Abstract

Detecting small objects in aerial imagery remains a formidable challenge due to their limited pixel resolution, scale variability, complex backgrounds, and inconsistent illumination conditions. To address these issues, we propose a novel hybrid object detection framework that synergistically integrates the real-time processing strengths of the YOLO architecture with the advanced hierarchical feature extraction capabilities of the Swin Transformer. The proposed YOLO-Swin hybrid model incorporates three key architectural innovations: (1) a Cross-Scale Feature Fusion Module (CSFFM) that effectively combines multi-resolution features from both convolutional neural network (CNN) and transformer-based pathways to enhance scale robustness; (2) a Context-Aware Small Object Enhancement Module (CASOEM) designed to enrich semantic representation and improve the detectability of small-scale targets; and (3) an Adaptive Anchor Assignment Strategy (AAAS) tailored to the spatial and statistical characteristics of aerial imagery. Extensive experimental evaluations conducted on widely used benchmark datasets—including DOTA, VisDrone, and FAIR1M—demonstrate that our model achieves state-of-the-art performance, outperforming baseline methods by achieving a 5.7% increase in mean Average Precision (mAP) for small object categories. Furthermore, the model maintains real-time inference capabilities, significantly reduces false negatives, and improves localization precision, particularly for objects smaller than 32×32 pixels. These results indicate the suitability of the proposed method for real-time aerial surveillance and remote sensing applications where precise small object detection is critical.