Research On Ship Classification In Remote Sensing Images Based On Heterogeneous Feature Knowledge Transfer

Synthetic aperture radar(SAR)has the advantages of wide coverage and all-weather earth observation,and is the main tools of ocean monitoring task.As the main carrier of marine activities,ships have naturally become an important object of marine monitoring.Therefore,ship classification in SAR images has become an important and critical part of marine monitoring,which helps managers understand the specific activities of ships and achieve comprehensive and effective marine monitoring and supervision.In recent years,in order to cope with the limited performance of supervised learning methods due to the lack of labeled data SAR images,methods based on transfer learning have become an emerging and promising research topic.However,existing related studies focus on homogeneous knowledge transfer from a source domain(single source),that is,the source domain and target domain data are required to have the same feature representation.Although these studies have achieved fruitful results,they are not suitable for heterogeneous scenarios where the source and target domains have different feature representations.As the heterogeneity difference between the source domains and the target domain has been paid more and more attention,how to effectively realize the heterogeneous knowledge transfer has become an urgent problem to be solved.Hence,in order to further improve the performance of SAR image ship classification,this dissertation focuses on heterogeneous knowledge transfer,and gradually conducts research on the heterogeneous transfer learning methods in single-source and multi-source situations.The main research contents and innovations of this dissertation are summarized below:1.In the single-source situation,a dynamic joint covariance alignment network is innovatively proposed.The network first utilizes feature projection networks to transform data into a common subspace to deal with feature heterogeneity,and then simultaneously performs classifier adaptation and joint distribution alignment to minimize domain bias,enabling heterogeneous feature knowledge transfer.Innovatively,the network considers using joint distribution alignment and dynamic weight assignment to learn the optimal common subspace,and the network is trained based on a semi-supervised learning setting,making the classification performance further improved by using unlabeled data.2.In the multi-source situation,a multi-source heterogeneous feature augmentation method is proposed.The method first uses feature augmentation projections to transform the data to the augmented common feature space to deal with the problem of feature heterogeneity,and then derives the formula based on the principle of the support vector machine(SVM)model to solve the optimal common space and data classification function,so as to achieve multisource heterogeneous feature knowledge transfer.Innovatively,the method deduces the formula of the multi-source heterogeneous feature augmentation method based on the SVM model,which is a multi-source domain extension of the original method;It solves the feature heterogeneity problem successfully that the multiple source domains are heterogeneous not only from the target one but also from each other.Overall,based on these research,following conclusions can be drawn: As for the feature heterogeneity differences between the source domain and the target domain,heterogeneous feature transfer is more helpful to improve the classification performance of the target domain than homogeneous feature transfer;There is complementarity between source domain knowledge,so under the same method framework,multi-source heterogeneous feature transfer can better improve target domain classification performance than single-source heterogeneous feature transfer.