Target Recognition Of SAR Image With Long-tail Data Distribution

Synthetic Aperture Radar(SAR)is widely used in military,agriculture,forestry and other fields,and it is an important means of ground monitoring.Target recognition in SAR images is of great value in research and practical application,and is currently a hotspot in the international SAR related research field.However,although the SAR image target recognition technology has made some achievements after years of research,it is still limited by the data distribution and application characteristics of SAR image in the complex actual combat environment.In practical application,due to the influence of uncertain environment,there is a lack of sufficient sampling rate and resident time to obtain a large number of samples of the corresponding target,and it is more difficult to obtain data of non-cooperative target.As a result,the data changes are relatively single and the number of non-cooperative target is small,which makes the data present a long-tail distribution.Due to the small number of tail samples and lack of diversity in the long-tail distribution,it is easy to lead to the problems of over-fitting and poor generalization performance of the model.This paper focuses on target recognition in SAR image with long-tail data distribution.First of all,this article in view of the SAR image data distribution imbalance easily lead to model fitting,poor generalization ability of the problem,put forward the decoupling of feature representation and classifier based on SAR image target recognition algorithm,the method study is divided into two stages,the representation learning and classifier learning,and put forward the balanced sampling method and the four different classifier learning strategies.The recognition accuracy and generalization ability of the model are improved by the method of balanced sampling and decoupling representation learning and classifier.Secondly,from the perspective of the "value" of SAR image labels information,this paper proposes the problem of long-tailed distribution of labels due to the unbalanced distribution of data.Aiming at this phenomenon,this paper proposes a SAR image target recognition algorithm based on contrastive learning.And the learning is divided into the contrastive learning stage and the classifier learning stage.In the SAR image target recognition method based on self-supervised contrastive learning,selfsupervised pre-training is first carried out,label information is discarded,features are extracted and stored more comprehensively,and then the first stage feature representation and balanced sampling are used for classifier training.This method improves the model label preference problem.In the SAR image target recognition algorithm based on weakly supervised contrastive learning,the category information of the label is introduced,so that the model can better extract the category features,and then perform the classifier training.Finally,the algorithm proposed in this paper is evaluated on the long-tail distributed SAR datasets,and different types of experiments are set for verification.The experiments show that the algorithm proposed in this paper achieves significant performance on the long-tail distributed SAR image target recognition task.