| Ships are the most important carriers for human activities at sea.Spaceborne synthetic aperture radar(SAR)occupies a dominant position in marine ship monitoring due to its all-day,all-weather,large-area,and longdistance observation capabilities.The use of remote sensing technology to monitor ships and accurately identify the type of ships is of great significance for maintaining national maritime security,maritime traffic management and maritime rescue.In the current research,the classification of ships based on SAR images still faces many problems to be solved.First,SAR ship classification is a finegrained classification problem,which is more difficult than ordinary image classification tasks.Fine-grained classification requires features to be more discriminative.However,limited by the imaging mechanism,SAR images can provide limited discriminative information.Therefore,for the distinguishing features of ships extraction is very important.On the other hand,fine-grained classification requires more precise measures to address the conundrum of interclass similarity and intra-class dissimilarity of ships.Second,SAR ship samples are difficult to label and the number is small,which has become the main bottleneck restricting the further improvement of the performance of supervised learning methods.This paper aims at the extreme case where the target domain(SAR domain)sample is completely unlabeled,which increases the difficulty of fine-grained classification of SAR ships.For the two problems,the main research contents and innovations of this paper are summarized as follows:This paper firstly constructs a high-resolution optical remote sensing ship classification dataset and a SAR ship classification dataset,which not only provides data support for the follow-up experiments of this study,but also contributes a publicly available high-quality dataset to the research in this field.Aiming at the problem of discriminative feature extraction of ships,this paper studies the influence of different backbone networks,the number of network branches,attention mechanism,and distance measurement on feature discrimination.It is found that by using different networks,increasing the network width,the embedding attention mechanism modules in the network and combining deep features with metric learning can improve the discriminativeness of extracted features.Optical remote sensing images have high resolution,high readability,easy acquisition and labeling,and belong to the same image data as SAR images,so they are more likely to share common features of ships with SAR images.For the case where the samples in the SAR domain are completely unlabeled,this paper uses the optical remote sensing ship image as the source domain to transfer the knowledge learned from the source domain to the target domain,and studies the feature alignment ability of the feature space shared by different unsupervised deep transfer networks.A deep transfer network that is most suitable for feature common attribute mining can be obtained.Based on the above studies,this paper proposes a new deep subdomain adaptation network DSAN++.The first half of the method is a dual-branch network embedded with an attention mechanism,which increases the discriminative ability of deep features from the perspective of feature extraction.Then,the local maximum mean discrepancy(LMMD)is used to measure the discrepancy between the highly discriminative ship feature distributions extracted by the dual-branch network in the common feature space,aligning the conditional distribution and marginal distribution of the cross-domain remote sensing data,and obtain a more class-discriminative common features.It effectively improves the performance of SAR ship classificationThe above methods have been systematically studied and discussed in depth on the two constructed ship datasets.Experiments show that the method proposed in this paper is superior to the existing unsupervised SAR ship classification methods and achieves the highest classification results,which effectively solves the problem of lack of SAR ship labeling samples and finegrained classification,and has certain reference value for future related research. |
