Design Of Land Battlefield Target Recognition System Based On Convolutional Neural Network

With the increasing complexity of modern warfare,it is more and more difficult to predict the real-time situation of war in the land battlefield,and it is necessary to master the trend of war development to win the victory of war.Therefore,it is necessary to find a method that can guarantee both real-time and accuracy to identify and feedback the results of different military targets in the land battlefield,so as to meet the operational requirements of the current land battlefield.Feature extraction is the most important step for target recognition in land battlefield.The traditional feature extraction method is easy to be influenced by illumination,deformation,occlusion and clipping,which leads to the low recognition rate of traditional methods.In this paper,the convolution neural network,which is superior to other methods in feature extraction,is selected,and the idea of transfer learning is combined with deep learning.A land battlefield target recognition system based on convolution neural network is designed,which can operate the extended dataset independently.The main work of this paper is as follows:Firstly,the structure and working principle of convolution neural network are deeply studied.Based on the characteristics of local perception and parameter sharing of convolution neural network,convolution neural network is introduced into the field of land battlefield target recognition.In this paper,a method for recognition of land battlefield targets is designed,which improves the accuracy of recognition compared with the traditional methods.Secondly,there are few SAR databases that have been published for research at home and abroad,so it is easy to use the original database as training set or test set to produce over-fitting problem.In this paper,a variety of data expansion methods are used to expand the exposed SAR image data set,so that the total amount of data can be expanded to 12 times of the original data.At the same time,through these data expansion methods,the recognition model fully solves the over-fitting and traditional identification methods are sensitive to the data translation,rotation and other changes.Thirdly,in order to solve the problem that traditional image learning and application scene are easy to migrate,this paper first studies the theory of migration learning,and uses image synthesis algorithm to synthesize 206 scene images containing targets.Secondly,a transfer learning model combined with convolution neural network is proposed,which improves the problem of local optimal solution caused by scene migration,and accelerates the training of model parameters.Finally,the United States MSTAR data set is used to test,and the comparison between the method combined with migration learning and other methods is given.The comparison results show that the method can identify at least three different military targets at the same time,and the recognition rate reaches to the level of at least three different military targets.Above,compared with other methods,the recognition rate is higher.Secondly,a set of integrated target background image synthesis and recognition image processing system is designed with MATLAB GUI graphical user interface,which can quickly and conveniently expand the sample size of the dataset and display the recognition results in real time,thus improving the work efficiency.It laid the foundation for the follow-up work.