ViDroneNet: An efficient detector specialized for target detection in aerial images

Abstract

In recent years, the widespread use of Unmanned Aerial Vehicles (UAV) has made UAV target recognition particularly critical. However, images captured by UAV are characterized by non-uniform object distribution, multi-scale changes, complex backgrounds, and flexible viewpoints, which is a great challenge for general object detectors based on common convolutional networks. To address these issues, we propose ViDroneNet (Vison Drone Network), an efficient framework specifically designed for target detection by UAV. Firstly, to overcome the challenges posed by multi-scale target, we design the Multi-Head Self-Attention darknet (MHSA-darknet) module and applied it to the backbone network. Then, for the problem of small target aggregation, we add a specialized probe head to deepen the understanding of the detailed information of dense small targets. Finally, we designed a Channel-space deformable convolution module (CSDC) and a new approach to feature fusion, both improved sensitivity to spatially distributed inhomogeneous targets and enhanced model robustness. Experimental results show that ViDroneNet outperforms state-of-the-art methods on the VisDrone and UAVDT datases, which were compared to achieve the highest mAP.