| Synthetic Aperture Radar is capable of round-the-clock operation in a variety of extreme environments.It has an irreplaceable role in the military,civilian and commercial fields.With the improvement of the technical level,the imaging technology of SAR tends to be mature,but there are problems such as difficulty in interpreting SAR image data.There is room for development of deep learning applied to SAR image interpretation research.The development of accurate and efficient SAR image interpretation technology is imminent.To address these issues,this thesis focusing on the recognition problem based on deep neural networks for small sample SAR image data.The main contents are: 1.In view of the speckle noise existing in SAR images,three filtering algorithms are selected to denoise the measured SAR images.The denoising work of the measured SAR image reduces the influence of noise on the SAR image target recognition,and prepares for the subsequent research on SAR image target recognition.2.In view of the restricted learning capability of shallow neural network and the problem of overfitting in training deep neural network with small sample SAR image data,a method of combining simulated SAR images with transfer learning is proposed to train Inception-Res Net-v2 network.The Inception-Res Net-v2 network introduces the concept of residuals,which gives the network a powerful feature learning capability and enables it to better acquire characteristics of SAR images.Meanwhile,the simulated SAR images with sufficient samples and similar characteristics are used to pre-train the network model until the model converges,and then the model is migrated.Secondly,the network is coached using the measured SAR images and the network parameters are simultaneously optimised to avoid overfitting and to improve the generalisation capability of the model.The ultimate implementation of 99.57% accuracy based on the MSTAR ten-class real-world dataset,and set up multiple sets of comparative experiments,it is verified that the deep neural network is better than the traditional method in the recognition research of small sample SAR image data.3.The problem of nasty loss of details in SAR images caused by training deep neural networks,an RCF(Res Net101-CBAM-FPN)neural network model is proposed.The model adopts Res Net101 neural network as the backbone of the network,and introduces Convolutional Attention Module(CBAM)and Feature Pyramid Model(FPN).The focus of the network model on the important features of SAR images is accomplished through CBAM,and the fusion of the underlying and high-level features of the network is achieved using FPN to decrease the characteristic damage of detailed image features and improve the recognition performance of the model.Secondly,combine the transfer learning method to complete the training of the network model.The final experiments demonstrate that the RCF network model has excellent recognition performance and is effective in suppressing the loss of detail features in the images.A recognition result of 99.60% was achieved based on the MSTAR dataset.The above has been verified through simulation experiments to show the high effectiveness and practicability of the approach offered in this article.It can provide reference value for the research and application of SAR image for target recognition. |
PDF Full Text Request