One-Stage Deep Channels Attention Network for Remote Sensing Images Object Detection

Jinyun Tang, Wenzhen Zhang, Guixian Zhang, Rongjiao Liang, Guangquan Lu

Abstract: Although existing remote sensing image object detection methods have made significant evolution in deep learning, they did not fully consider the problem of features loss caused by the correspondingly different importance of different channels of feature maps in the convolution pooling. Therefore, a one-stage deep channels attention network for remote sensing images object detection was proposed. First, through a multi-scale feature representation of the Single Shot MultiBox Detector (SSD) Network, the model can combine semantic information with detailed features to better integrate feature layers with different resolutions. Second, for each additional feature extraction layer, the squeeze and excitation (SE) module is introduced, which adaptively re-calibrates the interdependencies between deep channels, then they achieve the response of channel properties in order to learn more efficient feature information. According to experimental results on the RSOD dataset and NWPU VHR-10 dataset, the models proposed in this paper all realize advanced results and achieve state-of-the-art technical performance.