DHA-YOLOv8n: a lightweight network for detecting unmanned aerial vehicles

To address the challenges of low detection rates and high miss rates for small unmanned aerial vehicles (UAVs) in complex backgrounds during real-time detection tasks. In this paper, we introduce the DHA-YOLOv8n network. This newly proposed detection algorithm enhances detection performance while also decreasing both the parameter count and computational expenses of the network model. First, the small target detection layer P2 is introduced to reconstruct the YOLOv8n network structure to retain more shallow texture and edge features of the small targets to be detected in the input image; Secondly, the HetConv convolutional module is introduced, and the CSPDHet module is designed by combining the DualConv method to replace the C2f module in the backbone and neck structure of YOLOv8n. which reduces the model parameters and calculation costs. Finally, the downsampling convolution module ADown is introduced to obtain the image feature information and make the model lightweight to improve the deployment capability of edge devices. Compared with the baseline, the performance of mAP50 reaches 92%, and the accuracy rate reaches 96%, which is improved by 5.1% and 3% respectively. The memory occupied by the model is only 5.87MB. While fulfilling the deployment requirements of low memory consumption for edge devices, it simultaneously realizes highprecision real-time detection.