Target Classification Of Synthetic Aperture Radar Based On Semi-supervised And Unsupervised Learning

Synthetic aperture radar uses a small antenna to move along the trajectory of a long linear array at a constant speed and radiates coherent signals,and coherently processes the echoes received at different positions to obtain a higher-resolution imaging radar.It can obtain high-resolution radar images in extremely low-visibility environments,has all-weather,all-weather properties,and has a considerable ability to penetrate the ground.In the latest research,semi-supervised and unsupervised image classification methods are effective and very popular.This paper describes three semi-supervised and unsupervised image classification methods.In order to better fit the SAR image and reduce the computational complexity,this paper establishes a complex Gaussian Bayesian model in a dictionary learning environment to model the SAR image.Secondly,since the discriminant dictionary can be learned by modeling the distribution characteristics of SAR images,the discriminant dictionary of the distributed model must be learned.Finally,in order to solve the problem of limited labeled samples and the time consumption of existing algorithms,a semi-supervised online dictionary learning method is used to add training samples to update the dictionary.Experiments prove the performance of the proposed method by verifying semi-supervised characteristics,online dictionary learning characteristics,and classification effects.In addition,despite the great success of deep learning,there are still many limitations.The extent to which the computational properties of deep neural networks are similar to those of the human brain remains problematic.A special non-biological aspect of deep learning is the supervised training process using a back-propagation algorithm,which requires large amounts of labeled data and non-local learning rules for changing synaptic strength.This paper uses a learning algorithm that is not bothered by these two problems.It learns the weights of the underlying neural network in a completely unsupervised way.The whole algorithm uses local learning rules with conceptual biological rationality.The experiment proves the characteristics of the method by verifying the unsupervised characteristics and comparing the classification results of this method with other biological algorithms.Finally,due to the scarcity of SAR images,in order to make up for this shortcoming,a generation adversarial network is used to generate SAR images for training samples.This article uses an unsupervised method to learn the distribution of samples based on the WGAN method,and uses a small number of SAR images to generate sufficient samples to achieve the purpose of expanding the data set,and proves that the samples have been learned well by completing the classification task on the generated samples True distribution.