Anchor-Free Network for Multi-class Object Detection in Remote Sensing Images

The anchor-based object detection algorithms need many hyper-parameters artificial such as the threshold of intersection over union and the size of anchors, which may limit the detection performance to some extent. In order to better solve the problem of multi-class object detection in remote sensing images, this paper proposed an anchor-free object detection network that does not require any hyper-parameters artificial. The network improves the detection effect of small objects by using the improved feature pyramid network to fuse multi-scale feature maps more effectively, and Focal loss and loss of intersection over union (loll loss) are used as loss functions to optimize the network. The experimental results on the DOTA dataset show that the mean average precision (mAP) of this network is 71.02%, which is at least 10.5% higher than other existing networks. The anchor-free network proposed in this paper achieves superior detection performance in remote sensing images.