Abstract: Remote sensing images are complex in content and rich in categories, and there are numerous images difficult in discrimination, resulting in poor performance of remote sensing image retrieval. For this reason, the adaptive sample type discrimination (ASTD) method is proposed, which dynamically categorizes the sample types into simple samples, ordinary samples and difficult samples, and the network performs different degrees of learning based on the types of samples, so as to effectively improve the discriminative ability of features. Firstly, an SHash network is designed, which takes Swin Transformer as the backbone and adds a hash layer at the end of the network. This network captures the semantic information of images globally, improving feature representation and retrieval efficiency. Secondly, in order to make the same category of images more aggregated, and to better distinguish  different categories of images, a hash center is defined for each category. The center corresponding to the input sample’s own category is specified as the positive center of the sample, and the other centers are the negative centers of the sample. Finally, the sample type discriminative loss STDLoss is proposed to adaptively discriminate the type of samples based on distance relationship between samples and positive and negative centers, so as to improve the network’s ability to learn from each type of samples. Comparison with five hashing methods such as DSH, CSQ and SHC on two remote sensing datasets, UC-Merced and AID, experimental results show that the network trained based on the ASTD can better learn the features of the samples and improve the retrieval performance.

Key words: remote sensing image retrieval, sample type, Swin Transformer, hash, adaptive discrimination

摘要： 遥感图像内容复杂，类别丰富，存在较多难以判别的图像，导致遥感图像检索性能不佳。为此，提出自适应样本类型判别方法（ASTD），将样本类型动态地分为简单样本、普通样本和困难样本，网络依据样本的类型进行不同程度的学习，从而有效提高特征的判别能力。设计了一个SHash网络，该网络以Swin Transformer为骨干，在网络的最后加上哈希层，该网络能够在全局上捕获图像的语义信息，提高特征的表达能力和检索效率；为了让同一类别图像更加聚集，并更好地区分不同类别的图像，给每个类别定义一个哈希中心，规定输入样本自身类别所对应的中心为该样本的正中心，其他中心为该样本的负中心；提出样本类型判别损失STDLoss，根据样本与正负中心的距离关系自适应判别样本的类型，从而提高网络对各类型样本的学习能力。在UC-Merced和AID两个遥感数据集上与DSH、CSQ、SHC等五种哈希方法进行了比较，实验结果表明，基于ASTD方法训练的网络可以更好地学习样本的特征，提高检索性能。

关键词: 遥感图像检索, 样本类型, Swin Transformer, 哈希, 自适应判别