Research On Target Recognition Of High Resolution Remote Sensing Image Based On Convolutional Neural Network

High-resolution remote sensing imagery is a combination of various ground object information,and the feature details of the features are rich.The traditional artificial remote sensing image target recognition technology and pixel-based remote sensing image classification method can not meet the target recognition of high resolution remote sensing image.Precision requirements and timeliness requirements.Therefore,the research on new high-resolution remote sensing image target recognition methods has important theoretical research significance and application value.Deep learning is an important branch of machine learning,and it has performed well in practical applications and even reached the level of human beings.Convolutional Neural Network(CNN),as a deep learning model widely used in research and application,is considered to be the most powerful image recognition model.In this paper,the CNN deep learning model is applied to the automatic recognition of high resolution remote sensing image targets.A high-resolution remote sensing image scene classification framework based on CNN model is constructed,and a scene classification model based on AlexNet and VGG-Net convolutional neural networks is established.A regional convolutional neural network based on AdaBoost algorithm is proposed.An improved model that uses the AdaBoost algorithm instead of the selective search algorithm to extract candidate regions of the target to be identified from the image,and identifies the target features of the high resolution remote sensing image through the AlexNet convolutional neural network.The results of this paper are as follows:(1)Construct a high-resolution remote sensing image scene classification model based on convolutional neural network.In this paper,high-resolution remote sensing images Uc Merced and SIRI-WHU data sets are used as experimental data,and 1080 remote sensing images are selected as training sets.Firstly,the two proposed convolutional neural network models of AlexNet and VGG-Net are used to extract the scene features of the training set samples,and the extracted feature vectors are classified by using support vector machine(SVM).Then the model is verified and optimized.After the correction,the optimized scene classification model is obtained.Finally,the effectiveness of the convolutional neural network in the classification of high resolution remote sensing image scenes is verified by comparing the classification effects of the high-resolution remote sensing image scene classification method based on the word bag model.(2)An improved R-CNN model based on AdaBoost algorithm is proposed.In this paper,the DOTA dataset of the high-resolution remote sensing image dataset is used as the experimental data.Only the data samples containing the vehicle are retained,and the motor vehicles in the sample are used as the recognition targets.First,the AdaBoost algorithm is used instead of the Selective Search algorithm to perform candidate region extraction on the image.Convolution training of 1200 training data was then performed using the AlexNet convolutional neural network.Finally,the SVM is used to classify the features,and the candidate frame position is corrected to obtain the vehicle target recognition result.The experimental results show that the improved R-CNN model based on AdaBoost algorithm has higher recognition rate and faster recognition speed in high-resolution remote sensing image vehicle recognition than traditional R-CNN model.(3)The experimental verification of vehicle identification in the real-time dataset of high-resolution remote sensing imagery in Xuzhou City was established.This paper establishes the vehicle dataset of high-resolution remote sensing imagery in Xuzhou.The image of the dataset is from Google Earth.It cuts the high-resolution remote sensing image of Xuzhou City and Tongshan District,and eliminates the image of the vehicle.Only 380 images are retained.Remote sensing images of vehicles are included as experimental data sets.The size of each image is 227×227 pixels and the spatial resolution is 0.5 m.The R-CNN improved model based on AdaBoost algorithm is used to identify the real-time image.The experimental results show that the vehicle has good recognition effect.It proves that the improved CNN model has strong generalization ability in the field of high-resolution remote sensing image target recognition.