Research On SAR Automatic Target Recognition Based On Deep Neural Network

Synthetic aperture radar automatic target recognition(SAR ATR)technique has the ability of searching high-value targets accurately from large scene SAR images in a short time,and providing the category information for each target.It has become the hot issue domestically and abroad,which is of great importance in military and civil fields.This thesis focuses on the problem of SAR target recognition based on deep neural network,mainly contains SAR target recognition methods based on neural network ensemble and deep transfer learning.The main work is as follows:1.Aiming at the limited feature extraction ability of a single network,a method of SAR target recognition based on heterogeneous convolutional neural network ensemble is proposed.Several heterogeneous networks for feature learning of SAR images are constructed simultaneously to extract more complete classification and recognition information.It reduces the risk of network training falling into local optimum,and improves recognition accuracy and stability effectively.2.Aiming at the problem of network recognition ability declines with small-samples,the SAR target recognition method with fine-tuning of convolutional neural network is studied.Network initialization is realized via parameters of source domain pretrained model,which alleviates the low generalization ability of the network caused by insufficient training samples.The training efficiency and target recognition ability of the network are improved effectively.3.Aiming at the low transfer efficiency when the difference between source data and SAR images is large,a SAR target recognition method using domain-adversarial transfer learning is proposed on the basis of network fine-tuning.The parameters of the pretraining model are adjusted by domain-adversarial training,which enlarges the common feature space of source domain data and SAR target images,and further improves the SAR target recognition performance in the case of small training datasets.The work above is verified by simulation experiments.SAR target recognition is realized accurately and effectively.