SAR Image Target Recognition Based On Convolutional Dictionary Learning Model

As an active microwave imaging sensor,Synthetic Aperture Radar(SAR)can detect the ground at a long distance without the influence of weather and illumination.It has become one of the most important means in the field of earth observation.With the development of SAR imaging technology,a large number of high-quality SAR images can be obtained,which makes the SAR image target recognition technology more and more widely used in military,civil and other fields.The design of automatic target recognition system for SAR image has become a research hotspot at home and abroad.Based on the convolution dictionary learning model,this paper studies SAR image target recognition methods.The main contents of this paper can be summarized as follows:1.Aiming at the problems of dictionary simplicity and coding redundancy in traditional dictionary learning algorithms,this paper introduces multi-scale constraints and supervised information into the convolutional dictionary learning,and proposes a label-constrained multi-scale convolutional dictionary learning model,which can automatically extract geometric structure features of different scales in images.Label-constrained multi-scale convolution dictionary learning model(LMSCD)first extracts convolution dictionaries of different scales from images by using the method of convolution dictionary learning.When learning convolution dictionary,we add the category information to learn dictionaries,which are beneficial for the classification.Then,we convolute the convolution kernels of different scales with the original image to obtain the feature maps,which can be used as features for target recognition.The experimental results on MSTAR data sets show that the introduction of multi-scale constraints and supervised information can make the learned convolutional dictionaries more suitable for the target classification.The comparison with other methods shows that the proposed method has better recognition performance.2.Aiming at the problem that the existing SAR image target recognition methods do not make best use of the inherent physical characteristics of the target,this paper proposes a dual-stream feature fusion network based on the convolution neural network,attribute scattering center model and label-constrained multi-scale convolution dictionary learning model.The network can synthetically utilize the information of SAR image target both in the image domain and the physical mechanism.The two-stream fusion network consists of two branches: one branch combines with label-constrained multi-scale convolution dictionary learning model to interpret the image domain information of the target;the other branch combines with attribute scattering center model to interpret the physical information of the target;and finally,the high-level features of the two branches are fused.During the training process,two branches of the network are jointly optimized.The image domain information and the physical information of the target interact to ensure the optimal cost function.The experimental results on MSTAR real data sets show that the proposed network makes best use of the SAR image target information in the image domain and the physical mechanism,and has better recognition performance than the methods using only one kind of information.