A Research On Ship Classification Under The Conditions Of Few Labeled SAR Images

Synthetic Aperture Radar(SAR)plays a very important role in many civilian and military fields.Due to its all-time,all-weather,and large-scale observation capabilities,more and more applications using SAR images have appeared,including SAR ship image classification.Traditional SAR ship image classification algorithms tend to use a staged classification method.The first stage is to manually extract features,and the second stage is to input the features into a trained classifier,and finally get the classification result.However,this method relies too much on expert experience,and the feature scale of SAR images can be easily changed,which means the classification algorithm is not robust.Later,with the rise of deep learning technology,deep neural networks represented by convolutional neural networks can automatically complete feature extraction and classification tasks through end-to-end training,which improves algorithm accuracy and normalization capabilities to a certain extent.However,deep learning requires a large amount of labeled samples,and the cost of SAR image labeling is high,requiring more expert experience,so the classification accuracy of deep neural networks is limited.Therefore,this paper studies the SAR ship image classification algorithm under different few-shot conditions.First,based on a fully supervised learning scenario and a small number of labeled samples,an end-to-end training method based on stages is proposed.In the first stage,a deep convolutional neural network is designed,which can perform normal end-to-end image classification tasks,namely,automatic learning feature extraction process and feature classification process.The training process is the same as other image classification tasks.After the first stage of training is completed,the trained first stage network can be used as the backbone network,and some parameters of the feature extraction network can be fixed and independent,as an automatic feature extractor,as the preprocessing part of the second stage.In the second stage,a metric network is designed,which automatically clusters the features in the feature space,classifies them according to the clustering results,and obtains the final result.Secondly,based on the semi-supervised learning scenario,under the condition of fewer labeled samples and more unlabeled samples,an unlabeled data augmentation method is proposed.The data augmentation operation is performed on the unlabeled data and the data before and after the augmentation is measured in the network.The consistency of the output probability in the middle,to enable the network to learn the hidden information in it,so as to improve the classification accuracy of the algorithm.Finally,based on the transfer learning scenario,under the condition that the source domain needs to be transferred from the source domain to the target domain,and the target domain has a small amount of labeled samples,a network fine-tuning method based on adaptive features is proposed.The feature distribution in the target and source domain is used to adaptively adjust the feature value input into the classification network,thereby improving the classification accuracy of the algorithm.