Object Detection For UAV Images Based On Improved YOLOX
Yu Qiu, Yingying Feng, Haoran Chen, Dengyin Zhang · 2023
As the unmanned aerial vehicle (UAV) image has the characteristics of a complex background and dense targets with a small scale, a real-time detection algorithm YOLOX-RFS based on improved YOLOX-S is proposed in this paper. First, because a tiny target has few features that are easily lost in the deep network, the ConvNeXt Block module is introduced to enhance the representation ability of shallow neural networks and learn spatial details sufficiently. Then, the influence of the receptive field of the feature map on the performance of object detection for UAV images is analyzed. The receptive field of the low-level output feature map is reduced by changing the stem of YOLOX and adopting an inverted residual block. Finally, deformable convolution is used to compensate for the lost semantic information due to the decrease of the receptive field of the top-level feature map. Experimental results show that the algorithm in this paper achieves 38.8 % mAP and 45.5 FPS on the VisDrone2019 dataset. Without sacrificing the reasoning time, the mAP is improved by 3 % compared with the original algorithm, which is more suitable for object detection tasks in UAV scenarios.