Remote sensing images object detection based on channel aware attention networks

Due to remote sensing images have the characteristics of the large variations in object scale, complex background, and noise, object detection on remote sensing images has always been difficult problem in the field of computer vision. In order to make full use of the correlation of each channel of the feature map, our paper proposes a channel aware network based on FPN network and attention mechanism. The network adaptively assigns weights to the characteristics of each channel to achieve the purpose of suppressing background information and enhancing object information. In addition, our paper also combines the channel aware attention networks with FCOS network, and uses the GIOU Loss function and Focal loss function for training. Comparative experiments were conducted on the DOTA data set. The experimental results show that the FCOS algorithm based on the channel aware attention network has higher detection accuracy and detection speed, which is better than the existing object detection methods.