Research On Target Detection Algorithm In Optical Remote Sensing Image

Optical remote sensing image object detection is an important task in the field of computer vision and remote sensing image processing.At present,most of the remote sensing image object detection models based on deep learning are directly migrated from the general target detection model,which does not take into account the particularity of remote sensing images such as large-scale small targets,large background and indifferent directions,resulting in poor detection performance.On the other hand,there are many directional targets in remote sensing images,and the traditional horizontal labeling frame will also affect the detection accuracy.In order to improve the accuracy of remote sensing image object detection task,this paper improves the existing detection algorithm from two aspects: horizontal labeling and directed labeling.The main work of this paper is as follows:(1)On the horizontal frame object detection task,a remote sensing image object detection framework based on cyclic feature pyramid and cascaded double-branch detector is proposed.The cyclic feature pyramid is introduced to extract effective features for many times.Aiming at the inconsistency between regression and classification tasks,a double-branch model is used on the basis of cascade detection heads.In order to reduce the amount of calculation,the learnable anchor module is used when generating the region proposals,and the anchor frame is generated adaptively,which can also reduce the amount of subsequent calculation.Compared with the benchmark Cascade R-CNN,the accuracy of the algorithm on HRRSD data set is improved from 90.17% to 93.11%,an increase of 2.94%.The accuracy is improved from 16.09% to 25.80% on Vis Drone-DET dataset,which is 9.71%higher.(2)On the task of oriented frame object detection,a oriented remote sensing image object detection framework based on feature recombination and polarized attention is proposed.The feature recombination module is used to make the network pay more attention to the effective target area through weighting.A new oriented box labeling method is introduced to avoid the dislocation at the critical angle.The polarized attention module is used in the front end of the detector to improve the performance degradation caused by the inconsistent characteristics required for classification and regression tasks.Compared with the benchmark Rotated RPN,the accuracy of the algorithm is improved from 59.54% to 64.49%on Dior-R dataset,with an increase of 4.95%,and from 79.08% to 90.83% on HRSC2016 dataset,with an increase of 11.75%.(3)The two algorithms proposed in this paper are deployed on NVIDIA Jetson TX2 using HRRSD dataset and HRSC2016 dataset respectively,with good result,and the accuracy is only reduced by 0.55% and 0.17% compared with the server environment,which shows that the algorithm in this paper has practical use significance.