Abstract: In the object detection of aerial remote sensing images, aerial images were arranged in any direction under the overhead view, which had problems  such as large image size, arbitrary direction and complex background. In order to get better detection results in aerial images with complex backgrounds, we proposed a rotated object detection model in aerial images based on the attention mechanism. Firstly, RetinaNet was used as the baseline model, on the basis of  the original detector structure, additional  angle parameter was added to adapt to the  rotated object detection. Secondly, we proposed  a new channel semantic extracting (CSE)  attention module to  capture global semantic information and the channel relationship, and predicted the coarse bounding box and classification score.  Finally, the feature alignment and the improved Fast R-CNN detector head were used for fine processing to  further improve the detection accuracy and obtain the final classification and regression results. The experimental results show that the detection accuracy of this method on the public aerial remote sensing  DOTA dataset reaches 77.71%, which is superior to  other advanced rotated object detection methods.

Key words: object detection, aerial image, attention mechanism, deep learning