A Object Detection Network for Remote Sensing Images Based on Global Multiscale Attention and Local Context Dynamic Interaction

Abstract: Remote sensing image (RSI) object detection has been applied in the fields of marine information mining, building change detection, and ship monitoring. However, the application of advanced general-purpose object detection of RSI usually has significant degradation of performance. This is mainly due to the special natures of RSI such as complex background interference, arbitrarily oriented objects, small objects, and multi-scale. For these issues, we propose a novel object detection network (GLC-Net) for RSI based on global multiscale attention and local context dynamic interaction. Specifically, a global multiscale attention module (GMA) is designed to mitigate background interference. By incorporating the attention mechanism in the multiscale feature map, the model can focus on remote sensing object features and improve the efficiency and accuracy of detection tasks. To deal with the problem of the arbitrary orientation of objects, the rotated region proposal network is utilized with reference to the centroid offset representation. For the lack of information on small objects and the cross-scale problem, Contextual Dynamic Interaction Detection Head (CDI Head) is designed for feature filtering and enhancement by extracting local contextual information and ROI features in a one-to-one dynamic interaction. Extensive experiments had been conducted on two public benchmark datasets. The experiments demonstrate that the proposed GLC-Net has superior performance in RSI object detection.