Research On Remote Sensing Image Object Detection Based On Deep Learning

Object detection is a basic task in the field of remote sensing image analysis,and it is widely used in the fields of land resource monitoring,urban planning management,and military reconnaissance.In recent years,the remote sensing image object detection algorithms based on deep learning have made great progress.but there are still many problems to be solved urgently.This paper focuses on the difficulties in the field of remote sensing image object detection.The main research contents are as follows:In order to solve the problem that the detection results of single stage object detector in remote sensing image is not accurate,a cascade detection network CRNet was designed,which can obtain more accurate location results through two regressions.In addition,for the feature alignment problem in cascaded object detection,a feature alignment convolution based on deformable convolution is designed,which realizes the feature alignment between the convolution feature and the rotation bounding box.In order to solve the problem of inaccurate distribution of positive and negative samples in existing sample assignment method based on IOU（intersection over union）between the anchor box and the ground truth.A sample localization potential measurement metric based on the IOU between the regression bounding box and the ground truth is designed.According to the measurement metric,an adaptive sample assignment method is proposed to accurately allocate positive and negative samples.In object detection,the classification score is used to express the confidence of the output bounding box,but the classification score is not always a good estimate of the bounding box localization accuracy.To solve this problem,this paper proposes a IOU-aware classification score,which makes the classification score of the detection results consistent with the localization accuracy.Extensive experiments on the DOTA,HRSC2016 dataset show that the cascade detection network CRNet can greatly improve the detection accuracy of the object with large aspect ratio and densely arrangement.Adaptive sample assignment method can accurately divide positive and negative samples,thereby improving the detection accuracy of the network.The IOU-aware classification score can improve network performance without introduces additional calculations.Without bells and whistles,CRNet surpasses the current best algorithm on the HRSC2016 dataset with less calculation,reaching 90.26 m AP.After using multi-scale training and multi-scale testing,CRNet reached 79.42 m AP on the DOTA dataset,surpassing the complex two-stage detectors.It embodies the dual advantages of CRNet in efficiency and accuracy.