Research On Semi-Supervised Learning Method For SAR Image Target Recognition

Synthetic Aperture Radar is an active microwave imaging ground observation system using a coherent imaging mechanism.It transmits highly penetrating electromagnetic waves to the target that can penetrate clouds and vegetation,and quickly receives echo signals.Then use the Doppler effect to calculate the echo data to achieve long-distance high-resolution imaging,and obtain rich information such as the shape,structure,and scattering characteristics of the target.It has the advantages of all-day,all-weather,long-range and high-resolution imaging,and is widely used in the fields of national defense and security,survey and detection,environmental and energy monitoring.As the basic problem of SAR image analysis and interpretation,SAR image target recognition has received a lot of attention and research since the birth of SAR.Different from optical images,single-polarization SAR images have unique scattering mechanisms and speckle noise,which increase the difficulty of target recognition in SAR images.The traditional SAR image target recognition method is identified through two stages of feature extraction and classifier design.In the feature extraction stage,it is extremely dependent on prior knowledge and the steps are cumbersome.In recent years,end-to-end deep learning theories and methods have developed rapidly,and have obtained a large number of application verifications in the field of SAR image target recognition.However,deep learning methods need to rely on massive labeled data for feature extraction,and it is expensive and time-consuming to collect a large number of labeled samples suitable for deep learning methods in the field of SAR image recognition.In order to reduce the dependence of deep learning SAR image target recognition methods on labeled sample images,this thesis focuses on the hot topic of semi-supervised learning SAR image target recognition.Semi-supervised learning SAR image target recognition methods are proposed,such as multi-block interpolation hybrid operation,consistency regularization teacher-student model,dynamic threshold and weight adjustment,and early sample selection and retraining,to improve the generalization performance of the SAR image target recognition model in the case of a small number of labeled samples,the specific research contents and innovations of this thesis are as follows:(1)Aiming at the problem that the mixed image produced by the interpolation consistency algorithm of SAR image has local blur,the advantages and existing problems of the existing interpolation methods are analyzed,and the disadvantages brought by the interpolation are fully considered,and a multi-block mixed semi-supervised learning method for SAR image target recognition is proposed.Different from the global mixing idea of the existing interpolation mixing operation method,the primary and secondary mixed sample images are divided into different regions through multiple parallel ordered rectangular boxes.And randomly select a small area in each area,perform interpolation mixing processing,and generate a new sample whose visual effect is similar to the main mixed sample and contains some features of the secondary mixed sample.Use the new samples to train a convolutional neural network model through semi-supervised learning to improve the recognition and generalization performance of the model.The samples obtained after the interpolation operation of the proposed strategy is easier to be learned by the convolutional neural network model,which improves the target recognition accuracy of the semi-supervised learning SAR image.(2)For the current semi-supervised learning method of SAR image target recognition based on teacher-student model,there is a model training coupling effect,which affects the recognition accuracy of the training model.This thesis proposes a semi-supervised learning method for SAR image target recognition using a consistent regularization teacher-student model.Multiple network models are used to pseudo-label the unlabeled samples for each other for teachers and students to optimize the recognition performance of the network model.First,perform multiple data augment on unlabeled image samples,and use the student model to predict the samples after multiple augments in a consistent regularization manner.According to the predicted confidence probability of the unlabeled sample category based on the student model,the unlabeled samples are pseudo-labeled and sorted,so that the unlabeled samples are divided into consistent pseudo-labeled samples and confidence pseudo-labeled samples.Use the teacher model to re-label the divided confidence pseudo-labeled samples to solve the coupling problem existing in the semi-supervised learning method under the single teacherstudent model.The overall labeling accuracy of the student model for pseudo-labeled samples is improved to optimize the recognition performance of the model,and the proposed method is verified and analyzed through measured data.(3)For the semi-supervised learning method of SAR image target recognition in the early training process,there is a high error rate of pseudo-labeled samples,and the distribution of pseudo-labeled sample categories is not balanced,which affects the early training process of the model and thus affects the recognition accuracy of the network model.A dynamic threshold and weight adjustment method for target recognition in semi-supervised SAR images is proposed.The strategy of dynamic threshold and weight adjustment reduces the error rate of pseudo-labeled samples used to train the model and balances the category distribution of pseudo-labeled samples.During the training process,the number of pseudo-labeled samples of each category used to optimize the model is counted each time,and different thresholds are set according to the historical number of each category to balance the number of samples of each category.At the same time,the size of each sample loss item is obtained according to the calculation,and some samples are defined as "difficult to learn" samples,and a high learning weight is assigned to make the model pay more attention to the learning of "difficult to learn" samples and improve the recognition and generalization performance of the model.(4)In the case of a very small amount of labeled SAR image sample data,the semisupervised learning model has the risk of overfitting the labeled sample images,which makes the SAR image target recognition model have the problem of insufficient generalization performance,an early sample selection and retraining semi-supervised SAR image target recognition method is proposed.According to the principle that the early training model first learns simple and clean samples,the model set is obtained by initializing multiple models with different parameters and after multiple iterations of training,and the model set is used to pseudo-label the unlabeled sample data.Then,some pseudo-labeled samples with high confidence are selected from each category to form a low-noise pseudo-labeled sample set,which is combined with the labeled samples to train the model,thereby enriching the number of labeled samples.Finally,the low-noise pseudo-labeled samples and the pseudo-labeled samples labeled in real time by the model are further screened by the method of small loss selection,and the new semi-supervised learning SAR image target recognition model is retrained to obtain a higher recognition accuracy.