Research On Airport Recognition Based On Aerial Reconnaissance Imagery

With the development and application of high-end reconnaissance equipment such as drones and high-resolution satellites in recent years,more high-definition high-altitude aerial images have been obtained,and geospatial information and characteristic information of airport have been acquired through such high-resolution images.In the fields of people's livelihood,national defense,and economy,it is of great significance.However,traditional airport identification requires complex image preprocessing and feature selection and extraction steps.This will not only increase the amount of algorithm calculation,but also have a great influence on recognition accuracy.In traditional methods,relying on artificial characteristics does not have sufficient recognition capabilities for similar targets such as roads and bridges.In order to solve the above problems,this paper chooses convolutional neural network that can directly identify original image for comprehensive theoretical analysis and simulation verification,and combines deep network features with artificial features for research.The main work is as follows:?1?Analyze various classical convolutional neural network models,choose to transfer AlexNet network model parameters,modify the AlexNet model hierarchy for airport identification problems,change the last three layers on the basis of original network,and add a new full-connection layer.Aiming at disadvantage of ReLU function in Alex Net,use LeakyReLU instead of ReLU to improve the network performance.?2?Based on the network constructed in this paper,we train the network to achieve airport recognition on Matlab.Get 256?256?3 RGB images from Google Earth andexpand the dataset by flipping,mirroring,and other operations.Experimental result verify the excellent performance of the network constructed in this paper.?3?By analyzing misrecognized samples of the network constructed in this paper,it is mainly due to the fact that image contains a large number of straight-line features,and the background information is more prominent than airport itself,in order to improve the algorithm's ability to classify such objects.Select 10-dimensional texture feature based on GLCMs and employ PCA to reduct demensionality.Choose the first four dimension of PCA features to express texture feature of airport.The 2048-dimension output features of full-connected layer fc₁ of network constructed in this paper were extracted and fused with 4-D artificial features to generate fusion features.Use fusion features to train SVM classifier to recognize airport.The comparison results prove the availability of fusion features proposed in this paper.