Research On Ground Military Target Recognition Based On Deep Learning

At present,the civilian target recognition technology based on deep learning has been relatively perfect,but due to the security and confidentiality requirements of military systems and the complexity of military targets,the identification of ground military target weapon equipment types is still the focus in this field.Due to the modernization of the current equipment,during operations and drills,through different image acquisition equipment deployed on various types of equipment,you can obtain a large number of images in different scenarios and at different times,how to quickly and efficiently collect the large number of images collected The automatic identification of the system has very important tactical significance in future military exercises and reconnaissance activities in the battlefield environment.In recent years,the rapid development of deep learning technology has made it possible to identify the types of ground military targets based on collected images.This paper proposes an IRSE algorithm,a ground military target recognition algorithm based on deep learning.This algorithm is implemented by improving the original convolutional neural network model and the weight optimization algorithm.In order to complete the optimization of the network structure,several network models that are more commonly used for target recognition tasks are referenced,such as Inception V1,Inception V2,ResNet,VGGNet and Inception V2 as the basic network model,by adding SE after the basic network structure Module,redistribute the weights of the image features after the feature extraction of the convolution layer,and add the residual network module to learn the feature difference of the network to reduce the overall parameters of the network,and use the Adabound algorithm instead of SGD as the weight optimization in this paper algorithm.The experiment first uses the ImageNet data set for transfer learning,pre-trains the network structure,and then tests on private data set.The results show that,compared with several currently used target recognition network models such as VGG,Inception V1,and Inception V2,the improved network module is more suitable for target classification tasks on private ground military target data sets.Using this network model,Compared with the same network model before optimization,the recognition accuracy is improved by 1.5%,and the number of network convergence iterations is reduced by 10%.