<p>In drone-captured images, most of the objects to be detected are small. Moreover, the scale variations, dense distribution, and the inability of existing small object detection models to operate efficiently on drones make small object detection based on drone perspectives a challenging task. To improve the accuracy of small object detection, this paper proposes an improved YOLOv5-n model, ED-YOLO. First, an Efficient Edge Information Extraction Module (EStem) is designed to replace the initial two convolutional layers in the backbone network. This enhances the ability to capture edge information and multi-scale features from input images without increasing computational complexity. Second, a novel Multi-Path Coordinate Attention (MPCA) module is proposed, which, together with Deformable Convolution (DCNv2), reconstructs the C3 module, forming the proposed C3-DMP. This module improves small object detection performance through finer feature extraction and attention mechanisms. Next, to enhance multi-scale feature fusion while maintaining efficient feature extraction, a new Multi-Scale Efficient Decoupled Head (MSEhead) is designed, incorporating a multi-scale convolution module and decoupled head feature components. Finally, a detection head for small objects is added, and the detection head for large objects is removed to provide more refined features for small objects, thereby improving detection accuracy. Experimental results on VisDrone2019, TinyPerson and HIT-UAV datasets show that the proposed ED-YOLO outperforms other models. Compared with YOLOv5-n, it improves 12.7%, 7.8% and 1.9% on mAP@50, respectively, while the number of parameters decreases by 5%.</p>