Optical Remote Sensing Image Target Detection And Recognition Technology Based On Enhanced Region Convolutional Neural Network

With the improvement of the resolution of remote sensing images,the efficient and rapid detection and recognition of objects of interest in remote sensing images has become a hot research topic in remote sensing image interpretation.In recent years,the target detection algorithm based on deep learning has made breakthroughs in the field of target detection.Optical remote sensing images have certain similarities with natural scene images,so target detection method based on deep learning is also the target detection of optical remote sensing images.But due to the imaging methods and scales of remote sensing images are different from natural images,the application of deep learning methods in remote sensing image target detection still has the following challenges.(1)Small-scale targets are difficult to detect.In large-scale remote sensing images,small targets such as vehicles and airplanes are integrated with complex environments,and there are often a large number of false and missed alarms.(2)The target direction is randomly distributed.Due to the topographical characteristics of remote sensing images,the direction of the target is arbitrarily distributed;(3)the densely arranged target,a typical scene in the remote sensing image is that a large number of targets are distributed in the same scene,and how to accurately divide the boundary of adjacent targets is also an urgent problem to be solved.Aiming at the problems of the above-mentioned deep learning methods used in remote sensing image target detection,this paper has carried out a more detailed study,and verifies the effectiveness of the proposed algorithm by public remote sensing image datasets.The main contributions of this paper are as follows:Aiming at the problem of large scale difference and missed alarm of the small target in optical remote sensing image,a target detection method based on scale enhanced R-CNN is proposed.This method extracts features of different scales of different convolutional layers of backbone network.Small-scale targets use the characteristics of shallow convolutional networks,and large-scale targets use the characteristics of deep convolutional networks.In this paper,a series of experiments were carried out on the Google Earth image and the aircraft dataset in DOTA.The results show that compared with the traditional Faster R-CNN,the proposed method improves the detection performance of small targets obviously.Since the directions of the targets in the optical remote sensing image is arbitrarily distributed,this paper proposes an optical remote sensing image target detection method based on the rotating cascade network,which connects two rotation stages on the output of the original R2 CNN.By increasing the threshold for determining positive and negative samples,the distribution of positive and negative samples is changed,and the direction and position of the target are refined several times.This paper trains and verifies the performance of the network on the optical ship dataset HR-SHIP-15 collected from Google Earth.In the target recognition of the optical inshore ship of Zhongkexingtu Cup,19 kinds of ship targets were recognized in the United States and Japan.We obtain the second place,which further verified the performance of the algorithm.Aiming at the problem of arbitrary distribution and dense arrangement of target directions in remote sensing images,an optical remote sensing image target detection method based on region proposal rotation mapping is proposed.The intersection over union predictor is used to predict the positional confidence of the bounding box,and the positional confidence and the classification confidence are weighted together to serve as the basis for the non-maximum suppression class.Mapping horizontal candidate regions to rotation candidate regions with minimal increase through the region proposal rotation mapping module,and the scaling enhanced Ro I pooling is used to solve the alignment problem between the candidate region and the target.The performance of the algorithm is verified by the HR-SHIP-15 dataset and the typical dense arrangement category large-vehicle in DOTA.