BAIDet: remote sensing image object detector based on background and angle information

Remote sensing object detection, with large differences in object size, arbitrary orientation and tight arrangement, leads to difficulties in object recognition and localization. Therefore, a remote sensing image object Detector (BAIDet) based on Background and Angle Information is proposed in this paper. Firstly, a large convolutional kernel global attention module is designed to fully utilize the global information of remote sensing images by expanding the receptive field. And obtain the edge information of ground objects through deformable convolution. Secondly, an angle-sensitive probabilistic intersection-over-union loss function (AS-ProbIoU Loss) is developed for bounding box regression for oriented object detection. Finally, experimental results on four remote sensing image datasets, DOTA, HRSC 2016, UCAS-AOD, and DIOR-R, demonstrated the effectiveness of this method.