Full-Scale Aerial Target Recognition Method Based on Fully Convolutional One-Stage Object Detection

Convolutional neural network based detectors have been widely used in various target recognition tasks, and the ground target recognition on unmanned aerial vehicles (UAVs) platform has become a current research hotspot. However, there is still no suitable full-scale aerial target recognition method at present, which leads to low precision in various targets recognition on single UAV. Thus, a method based on Fully Convolutional One-Stage Object Detection (FCOS) is proposed to be employed in UAV target recognition to address this issue. The proposed method is an absolute one-stage detector which avoids using anchor. It deploys center-ness as the core of recognition in order to cope with targets at different scales. It takes Resnet as backbone and combined with FPN to extract features. The method is evaluated by a few experiments based on DOTA dataset with mAP value. The result illustrates that the method is capable of maintaining the accuracy of full-scale target recognition at 84 percent.