基于对抗深度学习的无人机航拍违建场地识别

Adversarial Deep Learning-based Unauthorized Construction Site Recognition Using UAV-assisted Aerial Photography

对违建场地的检测方法主要是通过人工对无人机航拍视频进行检查,存在检测精度低、识别性能差、工作效率低的问题。提出一种结合空间变换网络与Fast RCNN的生成对抗网络ASTN-Fast RCNN,通过深度学习与无人机航拍视频相结合自动识别检测处在建设初期的违建场地。将空间变换网络作为生成器,生成Fast RCNN目标检测器难以识别的旋转形变样本,并通过目标检测器与生成器的对抗式训练,提高检测器的鲁棒性。实验结果表明,该方法能够有效提高对无人机航拍违建场地的识别性能。

The detection method of unauthorized construction site is mainly to manually check the UAV-assisted aerial photography,leading to low detection accuracy,poor recognition performance and low work efficiency.This paper combines deep learning and UAV-assisted aerial photography for automatic recognition of unauthorized construction site in the early development stage,and proposes a generative adversarial network named ASTN-Fast RCNN,which combines Spatial Transformer Network(STN)and Fast RCNN.The STN(generator)is used to generate rotated samples that Fast R-CNN(detector)cannot easily recognize.Through such adversary training,the robustness of the detector is improved.The experimental results show that the proposed method can significantly improve the performance of unauthorized construction site recognition.