“In the field of remote sensing image object detection, researchers have proposed a new two-stage anchor free detection model based on Faster R-CNN, combined with anchor free detection and supervised mask attention techniques. This model guides the detection model to focus on the target area through attention mechanism and mask supervision method, improves the quality of target features, and adopts a dynamically adjusted soft label strategy to achieve reasonable label allocation and improve detection accuracy. On DOTA and HRSC2016 data sets, the average accuracy rate reached 76.36% and 90.51%, respectively, which exceeded most directional detection models, indicating the progressiveness and effectiveness of this method.”