DR-YOLO: An improved multi-scale small object detection model for drone aerial photography scenes based on YOLOv7

With the advancement of drone technology, detecting and recognizing ground targets from aerial perspectives has become crucial in various drone applications. However, object detection in drone imagery poses several challenges, including the prevalence of small targets, the significant impact of aerial perspectives, variations in target scales, complex backgrounds, and frequent occlusions. To address these issues, we propose DR-YOLO, a multi-scale target detection model specifically designed for aerial drone images, building upon the YOLOv7 framework. We introduce the Spatial Pyramid Pooling with Dilated Convolutions (SPPDSPC) module to enhance dense target feature extraction. Additionally, we incorporate a decoupled detection head tailored for small objects and redesign the number and sizes of detection heads. To handle complex backgrounds and varying target sizes, we embed the Multi-Scale Feature Fusion (HTLF) Module into the feature pyramid network, providing rich spatial information for detection heads of different scales. Furthermore, we utilize the Gaussian Wasserstein Distance (GWD) to refine the regression loss, leading to improved bounding box quality, faster convergence, and higher accuracy in small object detection. Experimental results on the VisDrone2019 dataset demonstrate a 14.8% increase in [email protected] and a 9.8% increase in [email protected] compared to the baseline YOLOv7, validating the effectiveness of DR-YOLO in detecting objects within aerial drone imagery. The code and results of our method are available at https://github.com/DRdairuiDR/DR-YOLO.