Research On Vehicle Target Recognition In SAR Image Based On Convolutional Neural Networks

Synthetic Aperture Radar(SAR)is a high-resolution imaging radar with all-day,all-weather and strong penetrating power.It plays an important role in national defense and civil fields such as military reconnaissance,urban planning,and environmental monitoring.With the continuous improvement of SAR image data acquisition capabilities and image resolution,how to interpret information in massive SAR image data has become a research hotspot.Convolutional Neural Networks(CNN),as a deep learning model,has algorithmically imitated the perception process of biological neurons,and achieved good results in the field of image target recognition.Research on CNN-based SAR image target recognition technology has certain engineering significance.Based on CNN,this paper focuses on vehicle target recognition under different conditions of SAR image data acquisition.The dataset adopted is MSTAR dataset,and the public part of the dataset contains SAR images of ten types of military vehicle targets.The main research contents of this paper are as follows:1)Aiming at the SAR image dataset with less data,which is easy to cause the problem of CNN model training overfitting,a data enhancement method based on Generative Adversarial Networks is designed.In this method,the full convolutional network is used as the generation model and the discriminant model of the Generative Adversarial Networks,and several optimization techniques are introduced into the model.The experimental results show that the SAR image data generated by the designed Generative Adversarial Networks model can obtain similar results to the real SAR image data on the five image quality evaluation indexes,that is,the acquired SAR image data can be used to supplement the SAR image dataset.2)Aiming at the difference in image characteristics between SAR images and optical images,the CNN model in the optical image field is directly applied to the target recognition of SAR images,and the result is not ideal.Based on the recognition model VGG16 in the optical image field,an improved CNN model is designed.The model takes into account the influence of the SAR imaging characteristics,the elevation angle of the imaging and the different variants of the vehicle target on the recognition performance.The optimization algorithms are introduced while the VGG16 structure is improved.The experimental results show that under the three recognition conditions,the designed CNN model can achieve a high recognition accuracy,and the recognition accuracy can be further improved when the training data is expanded to a certain extent.3)Aiming at the problem that the initial stability of the CNN model is poor due to random initialization of network parameters,which affects the recognition accuracy,the Transfer Learning theory is introduced into the designed CNN model.Part of the SAR image vehicle target data in the MSTAR dataset is taken as the training samples in the source domain.Model parameters trained in the source domain are transferred to the target domain model as the initial parameters of the network.The target domain training samples are used to fine-tune the network parameters to realize the information migration from the source domain to the target domain.The experimental results show that under the three recognition conditions,the designed CNN model can further improve the recognition accuracy after introducing Transfer Learning theory,and the convergence trend of the model is ideal.In order to verify the reliability of the algorithm,the experimental results are compared with seven excellent SAR image target recognition methods.The comparison results show that under the three recognition conditions,the SAR image vehicle target recognition method in this paper can achieve higher recognition accuracy.That is,it can promote the development of SAR image target recognition technology to a certain extent.