Aircraft Target Recognition And Classification Based On Deep Learning Network

Today in the 21 st century,China’s national economy continues to maintain steady growth,and mass travel generally seeks comfort and short-term travel,so aircrafts have become the preferred means of transportation for the masses.At the same time,combat aircrafts as a weapon has been widely used by the military.Whether used as a travel tool for civil aviation or in military applications,aircraft target classification has great application prospects.In recent years,deep learning has developed rapidly in the field of machine vision technology and achieved fruitful results.Aircraft image classification is a typical representative of fine-grained image classification.Its appearance,color,and strong interference factors of any modification seriously affect the accuracy of aircraft classification.This paper explores traditional classification methods,CNNs,and transfer learning strategies in this topic,mining hidden feature points in aircraft target images,building image classification models that are more suitable for aircraft features,and developing aircraft image recognition systems.This paper uses two aircraft target image datasets as the data base,and conducts experiments in this paper for the multi-class,high-accuracy challenge of fine-grained aircraft images.The main research contents and innovations are as follows:(1)Two aircraft image datasets DAP-6 and DWP-13 are created.There are 5702 images in the DAP-6 dataset,including civil,military and general aircraft.There are 18028 images in the DWP-13 dataset,mainly involving 13 types of fighters.Fully considering the importance of the dataset for the overall experiment,the two types of datasets created by the two types of data sets are widely used and highly compatible.The image samples include both images of the aircraft in flight in the air and aircraft on the ground.The image has certain promotion and practical value.In addition,Gamma correction is used to improve the brightness and activate edge features;the normalized gradient histogram,area mapping,and other operations are used to obtain the HOG vector of the image.(2)Combining HOG and SIFT with traditional SVM and KNN classifiers,based on four algorithms HOG + SVM,SIFT + SVM,HOG + KNN,and SIFT + KNN,respectively,to classify aircraft images.The experimental results show that HOG has more advantages than SIFT in aircraft image feature extraction,and the classification accuracy rate of HOG+SVM is 86.8%,which is higher than the classification accuracy rate of 73.8% achieved by HOG + KNN.The algorithm can effectively solve the problem of aircraft image classification.(3)An aircraft classification model based on DCNN is designed and built.Through several sets of comparative experiments,the network model is designed and the parameters are optimized.A classification model for aircraft targets is designed and the influence of different loss functions and optimizers on the model is analyzed.The mean square error loss function and the implementation of the SGD optimizer are determined.The classification effect is the best;then,a regularization cascade method is proposed,that is,a cascade method with BNlayer and Dropout of 0.5 to reduce overfitting and speed up the model convergence;the accuracy rate of the final designed DCNN aircraft classification model is 99.1%,which fully illustrates the effectiveness of the designed model.(4)Design and develop an aircraft target recognition system.Combining deep CNN model with transfer learning based on bottleneck layer feature extraction,using ResNet50,Mobile Net＿V2,Inception＿V3,and Inception＿ResNet＿V2 networks as feature extractors and identifying specific models through a layer of classification model,the solution is under DAP-6.The accuracy rate is over 99%,and the accuracy rate under DWP-13 is97.83%.The experimental results show that the method combining the deep CNN model and the bottleneck layer feature extraction strategy is effective for fighter type identification.In addition,the method combining the lightweight CNN model MobileNet＿V2 and the characteristics of the bottleneck layer has better real-time performance and requires less storage space.Based on this method,the aircraft target image recognition system is continuously developed,and as a result,high accuracy and low time-consuming targets are achieved.