Research On Object Detection In High-resolution Remote Sensing Image Based On Deep Learning

Object detection of high-resolution remote sensing images plays an important role in the interpretation and application of high-resolution remote sensing images.There are many applications in military and civilian applications such as disaster prediction,resource exploration,maritime fishery,and traffic supervision.Traditional object detection methods mainly use some traditional manual features(such as HOG,SIFT,Gabor,etc.),sliding windows and classifiers.The traditional manual features depend on the characteristics of professional domain knowledge and the data itself,and are computationally intensive and slow.They are more applicable to specific image data sets with smaller data volumes.With the increasing resolution of remote sensing images and increasing data volume,traditional manual feature methods have become increasingly difficult to meet the need for rapid development of remote sensing technology.Therefore,there is a need to find an algorithm that can quickly perform object detection from a large number of remote sensing image data sets.In recent years,deep learning,as a popular technology,can automatically extract features from image data,and has achieved great success in the field of natural language processing,speech recognition and computer vision.This has led to new developments in remote sensing image object detection.Based on the full research of remote sensing image object detection and deep learning related theories,this paper summarizes the main difficulties of remote sensing image object detection and the problems existing in current methods.It focuses on the research of the object detection method based on single convolutional neural network and integrated convolutional neural network for high-resolution remote sensing images.The main work is as follows:(1)For the traditional object detection method,sliding window is used to generate object proposal.At the same time,traditional manual features are used for object detection,which leads to low detection accuracy,object proposal redundancy,and lower efficiency.This paper proposes an object detection algorithm of high resolution remote sensing image which based on single convolutional neural networks.This algorithm first adopts multi-scale SLIC segmentation to replace the traditional sliding window method to obtain highquality object proposal.Then,the convolutional neural network is used toextract features from object proposals instead of the traditional manual features.At the same time,the Softmax classifier is used for subsequent classification.Experiments show that the method proposed in this paper has achieved a high recall rate and accuracy,and also has advantages in efficiency.(2)In order to make full use of the advantages of different network structures in extracting different levels of high-level semantic features,so that the performance of convolutional neural network based method is further improved,this paper proposes a remote-sensing image object detection algorithm based on multi-convolutional neural network.The method firstly selects a couple of convolutional neural networks with different structures and adjusts the structure of the network to high-resolution remote sensing image object detection.Then,a multi-depth convolutional neural network is used to extract semantic features of different levels of high-resolution remote sensing images.,while using multiscale segmentation to obtain the location information of the region to be classified;Finally,through the combination of location information and multilevel semantic features,using the trained network model to obtain the object detection results.