DCM-YOLOv8: An Improved YOLOv8-Based Small Target Detection Model for UAV Images

The development of Unmanned Aerial Vehicle (UAV) aerial vehicle detection technology holds paramount significance in the realms of traffic monitoring, management, and enforcement of traffic regulations. However, the high altitude view provided by UAVs results in smaller representations of vehicles in the images, and the limited resolution leads to blurred details, increasing the difficulty of recognizing small targets. To address this challenge, this study introduces DCM-YOLOv8, an innovative approach based on the YOLOv8n architecture. Firstly, the research introduces a Deep Attentional Spatial Pyramid Pooling module (DASPPF) that adeptly processes complex visual information by combining the large receptive fields of large convolutional kernels with the flexibility of deformable convolutions. Secondly, the paper details the design of a Cross-Level Multi-directional Weighted Pyramid Network (CLMiFPN), which is adept at integrating multi-scale information to improve detection performance for small-sized targets. Lastly, we present a novel MultiPath Coordinate Attention (MCA) mechanism, embedded within the dynamic head module, to replace the traditional detection head. This allows for more effective capture of information related to small targets in complex environments. Upon evaluation of the VisDrone2019DET dataset, the DCM-YOLOv8 model outperforms the baseline YOLOv8n, achieving a 13.2% increase in mean Average Precision (mAP) at a 50% Intersection Over Union (IOU) threshold. In addition, when compared to YOLOv8l, the proposed model registers a 2.1% improvement in mAP while requiring only 17.7% of the Giga Floating-point Operations Per Second (GFLOPs).