Multi-target Detection for Aerial Images Based on Fully Convolutional Networks
Haihong Chi, Xianjie Zhang, Xiangrui Gao · 2019
Multi-target detection for aerial images is widely applied for both military and civilian. A novel detection algorithm based on fully convolutional networks with fusion of multilayer convolutional features is proposed in this paper. This method predict directly the classification and bounding boxes of the objects in the aerial images based on pixel information, and improve the network framework by use of Group Normalizer and deformable convolution to promote accuracy of object detection. The experimental results on DOTA dataset with oriented bounding boxes show that the detection algorithm presented in this paper possesses fast detection speed and higher accuracy.