Research On Detection Algorithm Of Typical Objects In Remote Sensing Images Based On Deep Learning

The high-precision detection of typical objects in remote sensing images has important research significance and practical value in the military and civilian fields.At present,artificial intelligence and computer vision have achieved rapid development,and object detection in remote sensing images based on deep learning has received more and more attention.However,the high density of remote sensing objects,the wide range of scales,and the large changes in rotational attitudes make many challenges for high-precision remote sensing object detection.This paper studies the remote sensing object detection algorithm based on deep learning,analyzes the problems of high false alarms,large positioning errors,and missed detection of dense targets,and proposes improved algorithms.The main research work is as follows:Aiming at the problems of high false alarm and large positioning error in the Retina Net,this paper analyzes the fixed threshold positive and negative sample assignment method and the difference in sample importance in regression loss,and proposes based on adaptive threshold and soft sampled horizontal detection algorithm for remote sensing images.This paper proposes a sample suitability measure combining the center distance and the intersection over union,which solves the failure problem of the traditional intersection over union when measuring the closeness between the anchor box and the ground truth.This paper designs the adaptive threshold assignment strategy for positive and negative samples,which solves the problem of the insufficient number of positive samples for most objects in the case of single anchor boxes.This paper adopts the idea of soft sampling and proposes a method of re-weighting sample weights in regression loss by using sample suitability measure,which reduces the influence of poor suitability samples.Experimental results show that the remote sensing horizontal object detection algorithm based on the adaptive threshold assignment and soft sampling can significantly reduce false alarms,improve detection accuracy,and increase detection speed.Aiming at the problem of densely arranged oblique objects in a remote sensing complex scene,adjacent bounding boxes are mutually suppressed during NMS postprocessing,which leads to missed detection.This paper studies the multi-oriented object detection method in remote sensing images.Aiming at the problem of feature misalignment and large positioning error in the single-stage multi-oriented object detection algorithm,this paper proposes the multi-oriented object detection algorithm based on rotation feature alignment and refined regression.In this paper,the idea of convolution kernel deformation is adopted,and the four corner points of the initially predicted quadrilateral are converted into 3?3 feature sampling points aligned with the target through the proposed rotation feature alignment module.The refined regression network uses deformable convolution to extract features,and regression quadrilateral corner offset again to improve positioning accuracy.At the same time,this paper uses the anchor-free method to avoid a sharp increase in the number of anchor boxes during the multi-oriented object detection.Experimental results show that the multi-oriented object detection algorithm based on rotation feature alignment and refined regression has higher detection accuracy than other algorithms.