Research On Aircraft Classification Based On Deep Learning

With the continuous development of deep learning,the field of artificial intelligence continues to make breakthroughs and aircraft recognition systems play an increasingly important role in the military field.As an important part of the aircraft recognition system,the aircraft classification plays an important role in the identification of goals,the recognition of the enemy and so on.In the process of aircraft classification,the accuracy of aircraft image classification based on traditional methods is low due to various types of aircraft,large similarity between models,and severe texture interference.In response to these problems,the main work of this paper is as follows:Firstly,the original aircraft data set adopts multiple image expansion methods such as rotation,stochastic cropping,and Fancy PCA,and image enhancement methods,which effectively improves the diversity of aircraft data sets,and in turn enables the aircraft data set to satisfy the objective facts that deep learning requires large data samples.Secondly,because the labels in the data set of the aircraft have a multi-layer inclusion relationship,a Multi-Label Convolutional Neural Network(MLCNN)has been proposed to take advantage of the correlation between labels.MLCNN is to propose an improved multi-label structure based on Convolutional Neural Network(CNN).MLCNN can use the inclusion relationship between labels to sequentially set the sub-classifiers of each label at different layers of the convolutional neural network,to monitor feature extraction at different depths in the network layer.Then,considering that the training of a single MLCNN falls into the problem of local optimality and so on,combined with the idea of ensemble learning,a Bagging-based MLCNN integration model is proposed.The algorithm steps are: first,use a random sampling method to train multiple sub-MLCNN networks,and then multiple trained MLCNNs are used to test the classification of the aircraft,and finally the final labels of the aircraft are obtained through the voting method.Finally,using the Caffe deep learning framework,a single MLCNN and integrated MLCNNs are built and trained.It can be found through test set verification that the proposed method is optimal on the three labels: “manufacturer”,“family” and “variant”.Classification accuracy can reach 97.31%,93.11% and 87.44%.