Research On Optical Remote Sensing Target Detection Technology Based On Deep Learning

With the continuous development of remote sensing technology,remote sensing images are also used in various industries.In recent years,the resolution of optical remote sensing images has rapidly increased,and the amount of remote sensing image data has surged.However,the traditional remote sensing image target detection methods cannot meet the requirements in terms of accuracy and efficiency.At the same time,with the rapid development of the field of deep learning,in the field of natural images,target detection technology based on deep learning has made significant progress.Target detection is an important task in remote sensing image analysis.The target detection technology can quickly and accurately locate the region of interest in the remote sensing image.Traditional detection methods have the problems of low detection efficiency and poor detection accuracy for small targets.This paper studies A directional target detection algorithm for optical remote sensing images based on deep learning is introduced.The main work content is as follows:(1)The background and significance of the research on the target detection technology of optical remote sensing images are described,the difficulties and challenges of target detection in high-resolution optical remote sensing images relative to natural scene images are summarized,and the existing optical remote sensing image target detection technology is investigated Status and development.The theory of target detection technology based on deep learning is analyzed,and the commonly used convolutional neural network models and development trends are described.The advantages of deep learning in optical remote sensing image processing are pointed out,and it is determined to be an important development direction in the future.(2)Based on the original Faster RCNN model,an improved target detection model MFR-RCNN based on the rotation region proposal is proposed.Aiming at the problem of large changes in target scale and small targets in optical remote sensing images,the feature pyramid is introduced to make full use of the multi-layer information of the backbone network to detect targets on the multi-layer feature map and detect small targets on the large feature map,Detection of larger targets on small feature maps effectively improves the accuracy of target detection.In addition,due to the angle of view,the objects in the optical remote sensing image have directional diversity and the accuracy of the horizontal detection frame is reduced in crowded scenes.A rotating area suggestion network is proposed to realize the orientation detection of the optical remote sensing image.Experiments show that the proposed improved method obviously improves the detection accuracy,the directional detection frame is more suitable for the actual needs of optical remote sensing images,and the detection performance on the published optical remote sensing image data sets is good.(3)The RPN network needs to calculate the IoU between each Anchor and the real target frame to divide the positive and negative samples.Due to the large number of candidate frames,the RPN network calculation in the training phase is too large.In response to this problem,SA-RRPN is proposed based on the idea of Anchor Free,which directly generates candidate frames based on the points on the feature map,without the need for correction by Anchor,which effectively reduces the amount of RPN network calculations and eliminates the need to manually design Anchor It is difficult,and the problem of sample imbalance is effectively improved by Focal Loss.While improving the detection speed and reducing the occupation of training video memory,the accuracy has also been improved to some extent,which proves that the method is feasible.