摘要： 哈希方法由于低存储、高效率的特性而被广泛应用于遥感图像检索领域。面向遥感图像检索任务的无监督哈希方法存在伪标签不可靠、图像对的训练权重相同以及图像检索精度较低等问题，为此，提出一种基于深度多相似性哈希（DMSH）的遥感图像检索方法。针对优化伪标签和训练关注度分别构建自适应伪标签模块（APLM）和成对结构信息模块（PSIM）。APLM采用K最近邻和核相似度来评估图像间的相似关系，实现伪标签的初始生成和在线校正。PSIM将图像对的多尺度结构相似度映射为训练关注度，为其分配不同的训练权重从而优化深度哈希学习。DMSH通过Swin Transformer骨干网络提取图像的高维特征，将基于语义相似矩阵的伪标签作为监督信息以训练深度网络，同时网络在两个基于不同相似度设计的模块上实现交替优化，充分挖掘图像间的多种相似信息进而生成具有高辨识力的哈希编码，实现遥感图像的高精度检索。实验结果表明，DMSH在EuroSAT和PatternNet数据集上的平均精度均值较对比方法分别提高0.8%~3.0%和9.8%~12.5%，其可以在遥感图像检索任务中取得更高的准确率。

关键词: 深度无监督学习, 遥感图像检索, 特征提取, 哈希学习, 伪标签

Abstract: Hashing methods are widely used in remote sensing image retrieval owing to their low storage and high efficiency.Unsupervised hashing methods for remote sensing image retrieval tasks are often associated with unreliable pseudo-labeling, the same training weights of image pairs, and the low accuracy of image retrieval.Hence, a remote sensing image retrieval method based on Deep Multi-Similarity Hashing (DMSH) is proposed herein.The Adaptive Pseudo-Labeling Module (APLM) and Paired Structure Information Module (PSIM) are established to achieve optimal pseudo-labeling and training attention, respectively.The APLM uses the K-Nearest Neighbor (KNN) and kernel similarity to evaluate the similarity relationship between images for the initial generation and online correction of pseudo-labeling.The PSIM maps the multi-scale structural similarity of image pairs to training concerns and assigns different training weights to optimize deep hash learning.The DMSH uses the Swin Transformer backbone network to extract the high-dimensional features of images as well as a pseudo-labeling based on a semantic similarity matrix as supervised information to train the deep network.Furthermore, the network is alternately optimized on two modules designed based on different similarity degrees to fully exploit the multiple similarity information among images and then generate highly discriminative hash codes to retrieve remote sensing images with high precision.Experimental results show that the mean Average Precision (mAP) of the DMSH improved by 0.8%-3.0% and 9.8%-12.5% on the EuroSAT and PatternNet datasets, respectively, compared with the optimal results yielded by other methods analyzed.Hence, the proposed method can effectively improve the retrieval accuracy of remote sensing images.

Key words: deep unsupervised learning, remote sensing image retrieval, feature extraction, hash learning, pseudo-labeling

中图分类号: 

• TP391