Target Detection Method Of Remote Sensing Image Based On Deep Learning

The application of deep learning to target detection is one of the key research contents in the field of machine learning.The research background of this thesis is target detection of remote sensing image.Aiming at the problems of low detection accuracy and poor real-time performance of existing remote sensing image target detection algorithms,the solution based on deep learning is analyzed and studied.This solution has made significant progress in both detection accuracy and detection speed compared with previous traditional algorithms.(1)Traditional image processing algorithms,including image preprocessing algorithm and feature extraction algorithm,are analyzed and studied in this thesis.In the image preprocessing algorithm,image enhancement algorithm and image segmentation algorithm are mainly analyzed and studied.In the image enhancement algorithm,the median filtering algorithm and the guided filtering algorithm are emphatically analyzed.In the image segmentation algorithm,the most traditional threshold-based image segmentation algorithm is analyzed,and the clustering-based image segmentation algorithm is emphatically studied.Finally,an image segmentation algorithm based on depth learning is introduced,and the research focus of this subject is located in the field of deep learning.In addition,in view of the inherent characteristics of remote sensing images,such as cloudy fog,this subject focuses on the analysis and research of defogging algorithm.In the feature extraction algorithm,firstly,the Scale-invariant Feature Transform,Histogram of Oriented Gradient,and Local Binary Patterns are studied in this subject.Secondly,the classification algorithm,Support Vector Machine,is emphatically studied.This subject integrates HOG features and SVM,and designs a traditional target recognition model based on HOG + SVM.(2)The research focus of this subject is target detection based on in deep learning.In this subject,Region-CNN and its derivative model,You Only Look Once algorithm and its derivative model,and Single Shot Multibox Detector(SSD)algorithm are analyzed and studied.The object detection of remote sensing image based on SSD algorithm is mainly studied.According to the training results of data sets and original SSD algorithm,two improvements are proposed in this subject: one is to get more rich image feature information by adding feature extraction layer;the other is to get more fine-grained candidate regions by improving the criteria of candidate box delimitation.Through the improvement of these two algorithms,the improved model proposed in this subject improves the detection accuracy on the premise of real-time detection.(3)In the last part of this subject,the effectiveness of the improved algorithm is verified by designing comparative experiments.Contrast experiments mainly include traditional target detection algorithm,Faster-RCNN target detection algorithm,YOLOv1 target detection algorithm,original SSD target detection algorithm and the improved SSD algorithm proposed by this subject.By comparing the detection accuracy,detection speed,small target detection effect and so on,the validity of the improved algorithm proposed in this subject for the remote sensing image dataset of the marine ship is verified.