Aircraft Image Recognition Based On Deep Convolution Neural Network

In the modern battlefield environment,the enemy target detection and identification has always been a very important and realistic problem.The traditional target recognition program mainly relies on radar system,millimeter wave,laser and other means to distinguish between enemy and target,which are susceptible to electromagnetic and cost a lot as well.In the recent years,with the maturity of the deep learning theory,a variety of applications based on depth learning has coned one after another in the fields of the computer vision,speech recognition,natural language processing.Convolution neural network have a very wide range of applications.in the object classification,image segmentation and other issues as one of the most effective means in the field of computer vision.To this end,this paper carried out the aircraft image recognition algorithm research based on the deep neural networkBased on TensorFlow,this paper studies the convolution neural network for aircraft identification.The algorithm does not need to separate the target object from the background,and the target recognition task of the aircraft is divided into two parts:the candidate region generation and the aircraft classification.As for the problem of candidate region generation,this paper draws on the idea of faster-rcnn to construct a candidate region extraction network,using the convolution neural network to optimize the region position and size of the candidate frame and obtain the classification probability by generating the candidate box in batches of each image.In addition,this paper designs a candidate region selection algorithm based on hierarchical clustering,which greatly reduces the test run time by preliminarily screening candidate regions.The part of aircraft classification extracts several common types of aircraft targets to build an image classification neural network based on the AleNet.And use the method of migration learning to solve the problem of difficult training of network labels with scarce labels,as well as further improve the classification effect through the data enhancement,batch normalization.Finally,the training and experimental verification of aircraft target recognition network based on deep learning are carried out.This paper captures the five types of aircraft images in the ImageNet image database through web crawlers,marks the aircraft types and candidate area for all types of images,and enhance data through the methods of flip,contrast conversion and others,and then dived it into train set and test set after sort it out of order.Experiments show that the classification accuracy of this paper is more than 90%,and the choice of aircraft position is basically in line with the actual situation,and the running speed is greatly improved compared with the traditional algorithm.