Research On Detection And Recognition Methods Of High-resolution Remote Sensing Image

With the rapid development of deep learning field,high-performance object detection technology has become more and more mature.In the task of target detection,target detection for high-resolution remote sensing images is one of the difficult problems,and it also has important research significance and practical application value.Remote sensing images are usually obtained by aerial photography equipment carried by drones or satellites,and have practical application value in many aspects such as unmanned driving,military missions,and geological exploration.With the improvement of the resolution of remote sensing data,the information provided by the image becomes more and more abundant,which puts forward higher technical requirements for improving the ability of automatic image interpretation and the information utilization of remote sensing data.Compared with the target detection of general objects,remote sensing targets usually lack sufficient appearance information,and it is difficult to accurately detect them through detailed information such as appearance scale.Although significant breakthroughs have been made in target detection algorithms driven by deep learning,the detection of remote sensing targets is still unsatisfactory.On the high-resolution remote sensing detection dataset DOTA,there is a significant gap in the detection performance of different categories of objects,such as small vehicles and bridges,which have large differences in scale.It can be seen that remote sensing target detection is still full of challenges.The related methods in recent years are mainly based on the general object detector,solve the scale difference problem through the feature pyramid network,and add an angle regression branch to solve the target rotation frame problem.Existing remote sensing object detectors can also be divided into one-stage and two-stage approaches.Some accuracy-focused two-stage detectors,such as ICN,Ro I Transformer,and SCR Det,show decent performance in detecting dense small objects.Compared with the above two-stage detectors,R3 Det and RS Det are based on single-stage detection methods,which pay more attention to the trade-off between accuracy and speed.Based on the two-stage detector,this paper designs a more accurate detection method that can realize remote sensing target detection and recognition.To solve the problem of high-resolution remote sensing image detection and recognition,we propose a remote sensing object detector based on instance segmentation correction.For the dense distribution of remote sensing images and small target objects,the regression branch of dense local regression is adopted.And adopt the discriminative Ro I Pooling scheme to realize adaptive weighting to enhance the discriminative features.In addition,the instance segmentation technique is used to extract more accurate target appearance contours,so as to obtain detection frames that are more suitable for the target.The method can accurately locate and classify objects without the need for angle regression or complex bounding box offset calculations,and the problem of rotated object detection can be solved using mask information.To verify the effectiveness of the method,we conduct experiments on the DOTA v1.0 dataset of Wuhan University and the NWPU VHR-10 dataset of Northwestern Polytechnical University,respectively.In the method comparison experiment on the DOTA dataset,we achieved an average precision(m AP)of 76.46%,which is 3.18%higher than the existing remote sensing target detection method in the overall evaluation result.On the backbone network,we performed a set of ablation Experiments show that our method is scalable to a certain extent;78.6% m AP is achieved on the NWPU VHR-10 dataset.In addition,we conduct ablation experiments in the regression and segmentation branches,which proves that our method can only solve remote sensing.Effectiveness on object detection problems.