Enhanced Small-Object Detection in UAV Images Using Modified YOLOv5 Model

This study presents a modified YOLOv5 algorithm specifically designed to enhance small-object detection in unmanned aerial vehicle (UAV) images. Traditional object detection in UAV images is particularly challenging due to the high altitude of the cameras, which results in small object sizes and varying viewing angles. To address these challenges, the algorithm incorporates an additional prediction head to detect objects across a wide range of scales, a channel feature fusion with involution (CFFI) block to minimize information loss, a convolutional block attention module (CBAM) to highlight the crucial spatial and channel features, and a C3 structure with a Transformer block (C3TR) to capture contextual information. The algorithm additionally employs soft non-maximum suppression to enhance the bounding box scoring of overlapping objects in dense scenes. Extensive experiments were conducted on the VisDrone-DET2019 dataset, which demonstrated the effectiveness of the proposed algorithm. The results showed improvements with precision scores of 55.0%, recall scores of 44.6%, mean average precision scores of mAP50 = 50.9% and mAP50:95 = 33.0% on the VisDrone-DET2019 validation set, and precision of 50.8%, recall of 37.3%, mAP50 = 44.2%, and mAP50:95 = 27.3% on the VisDrone-DET2019 testing set. The improved performance is due to the incorporation of attention mechanisms, which allow the proposed model to stay lightweight while still extracting the features needed to detect small objects.