Research On Detection In Remote Sensing Images By Depth Learning

In recent years,with the rapid development and improvement of aerospace remote sensing technology and the image sensor technology,the number of high-resolution images has remarkably increased.The high-resolution images bring more information as well as more complex background information.How to quickly and accurately obtain the interested target positions of aircraft and ships from high resolution remote sensing images has become one of the important research directions in the field of remote sensing satellite image analysis.Due to the large amount of interference in optical remote sensing images,it is very difficult to rely on the traditional image methods to ensure the accuracy of recognition in complex background and illumination changes.At the same time,with the continuous development of deep learning,deep learning has made remarkable achievements in many fields such as the classification,target detection and segmentation in daily images.This paper is to use deep learning method to solve the task of detection and segmentation of targets in optical remote sensing images.The research work and results are as follows:(1)Taking the airplane as an example,Single Shot Multi Box Detector algorithm based on depth learning realizes the target detection in optical remote sensing images.Single Shot Multi Box Detector algorithm is a one-stage depth learning target detection algorithm for daily image design,which is fast and accurate.The paper analyzes the problems caused by the differences between the optical remote sensing images and the daily images,such as: insufficient data volume,relatively small target,and the image data obtained as rectangle.New algorithm can be used in optical remote sensing image and can accurately detect the aircraft target.(2)A new algorithm framework is designed to combine the detection targets in deep learning with the saliency segmentation in traditional image technology,and realize the weak supervision segmentation of aircraft targets in optical remote sensing images.The deep learning algorithm performs well in image segmentation because it is based on a large number of segmentation data.In the face of the problem that no labelling aircraft segmentation data and too hard for manual labelling,this paper combines the Single Shot Multi Box Detector algorithm with the saliency segmentation algorithm.Only using the data of the external rectangular box of the aircraft in the training set successfully and Automately obtained the segmentation results of the aircraft in the test set.(3)A new deep learning algorithm,Faster R-CNN-R,is presented to detect targets in optical remote sensing images.This paper Takes the boat as an example.At present,all the end-to-end deep learning algorithms can only output external rectangular boxes.But because of the large size of the ship,the angle of the remote sensing image is arbitrary,it is difficult to represent the real location of the external rectangle box.We added the Faster R-CNN algorithm to the angle information and succeeded Rotating rectangular box.With the speed basically invariable,the rotation frame AP has obvious promotion.Based on the above research direction,we conducted a large number of experiments based on the Mxnet depth learning framework on the Ubuntu system.The interface was created based on python and Tkinter.Experiments proved that the algorithms are better than traditional algorithms of image technology and the validity.The algorithms can be applied to the actual optical remote sensing image segmentation and detection detection task.