SAR Image Target Recognition Based On Semi-supervised Learning

Synthetic aperture radar(SAR)is widely used in military security and civil fields because of its advantages such as all-weather and all-weather.The amount of SAR image data is huge and the cost of manual labeling is high.Commonly used SAR image target recognition methods and deep networks that require SAR image data tags for supervised learning are difficult to perform high-accuracy target recognition.In view of the above problems,this paper proposes to use semi-supervised learning method to recognize SAR image targets.Semi-supervised learning uses a small amount of labeled SAR image data to solve the problem of target recognition on a large number of unlabeled SAR image data.The main research content of this paper has the following three aspects:1.Use the 3D block matching algorithm(Block-matching and 3D filtering,BM3D)to reduce the noise of the SAR image.Experiments were performed with the 2S1 data samples in the MSTAR data set,and the peak signal-to-noise ratio(PSNR)was increased to 32.7196 by using the BM3 D algorithm to reduce noise,and the structural similarity(Structural similarity,SSIM)value was 0.74283.Close to 1.2.Propose a semi-supervised ladder network(Semi-supervised ladder network,SSLN)SAR image for target recognition.The semi-supervised ladder network inputs the SAR image labeled samples into the encoder training classification function of the semi-supervised ladder network,and a large number of unlabeled samples are transmitted to the encoder for training the network,and the network performance is optimized by minimizing the total target loss function of the network.Experiments were carried out with the MSTAR data set,combined with the BM3 D noise reduction algorithm,compared with the previous method,the recognition accuracy has been improved and the running time has also been reduced.3.Propose a SAR image target recognition based on a semi-supervised variational autoencoder network(Semi-supervised variational autoencoder,SSVAE).Use the generation model of the latent feature discriminant model to learn new latent representations of SAR images,use the latent variable embedding learning of the original SAR image to generate a semi-supervised model,continuously optimize the model parameters through the deep network,and infer the loss function and missing labels of the labeled samples by variational inference The lower limit of the sum of loss functions of the samples reaches the optimization of target recognition.Experiments with the MSTAR data set,combined with the BM3 D noise reduction algorithm,the recognition rate is increased to 96.37%.Compared with the semi-supervised ladder network,the semi-supervised variational autoencoder runs more time than the semisupervised ladder network as the pixels of the target image become larger,and the stability of the semi-supervised variational autoencoder is not as stable as the semisupervised ladder network.