面向遥感图像检索的无监督哈希融合方法  

Unsupervised Hash Fusion Method for Remote Sensing Image Retrieval

作　　者：江文聪 王思佳 葛芸[1] JIANG Wen-cong;WANG Si-jia;GE Yun(School of Software Engineering,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学软件学院,南昌330063

出　　处：《南昌航空大学学报（自然科学版）》2024年第1期43-51,共9页Journal of Nanchang Hangkong University(Natural Sciences)

基　　金：国家自然科学基金(42261070,41801288)。

摘　　要：遥感图像数据规模庞大,且大部分数据没有标注,因此能够独立于数据标注的无监督哈希算法更适用于遥感图像检索。本文提出了一种针对无监督哈希的融合方法,首先,通过随机游走算法得到遥感图像预训练特征之间的流形相似度,并结合特征之间的余弦相似度构造相似指示矩阵,该矩阵可以度量无监督哈希码的有效性。然后,将哈希码的有效性作为节点,哈希码之间的排序相关性作为边来动态地构建关联图,并将图中的连通分量作为哈希码的组合,避免可能产生退化结果的哈希码组合,进而降低计算复杂度。最后,将哈希码的归一化有效性作为权重,对每种组合方案进行自适应的晚期融合,生成判别能力更强的哈希码。在2个数据集上的一系列实验表明,该方法能自适应地选择出合适的融合方案,有效提升融合哈希码的检索性能,并且在付出更小训练代价的情况下,获得接近穷举方法检索性能的融合方案。Because the scale of remote sensing image data is enormous and most of the data is not annotated,the unsupervised hashing algorithms that independent of data annotation are more suitable for remote sensing image retrieval.In this paper,a fusion method for unsupervised hashing was proposed.First,the manifold similarity between pre-trained features of remote sensing images was obtained by the random walk algorithm,and the similarity indication matrix was constructed by combining the cosine similarity between features,which could measure the validity of the unsupervised hash codes.Then,the validity of the hash code was used as a node,and the sorting correlation between hash codes was used as an edge to dynamically construct an association graph.The connected components in the graph would be adopted as the combination of hash codes,avoiding those hash code combinations that may result in degradation and reducing the computational complexity.Finally,using the normalized validity of the hash codes as a weight,adaptive late fusion was carried out with each combination scheme,which could promote the hash codes with stronger discrimination ability generating.A series of experiments on two datasets show that the proposed method could adaptively select suitable fusion schemes,effectively improve the retrieval performance of fused hash codes,and obtain fusion schemes with retrieval performance close to that of exhaustive methods at a lower training cost.

关 键 词：遥感图像检索 无监督哈希 自适应融合 流形相似性 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]