ABSTRACT

The advancement in the field of remote sensing (RS) has offered a vast number of RS images with higher resolution. Nowadays, remote sensing image retrieval (RSIR) has become a challenging task for researchers due to the complicated contents and specific characteristics of RS images. Content-Based Image Retrieval (CBIR) methods create powerful tools for mining large RS image databases. Content-based RSIR aims to acquire the images with similar visual content based on a query given from a large-scale RS image library. Most of the previous works used pre-trained convolutional neural network (CNN) to form a scene illustration for the classification of RS scenes. In this work, an effective RSIR using hybrid VGGNet (Visual Geometry Group Network) CNN with red deer algorithm (RDA) is presented for the appropriate retrieval of RS images based on the query image. The proposed hybrid VGGNet CNN model integrates dimensionality reduction (DR), feature extraction (FE), loss function optimization, matching process and relevance feedback (RF) mechanism. The technique used to reduce the dimensionality is Principal Component Analysis (PCA). Next, feature extraction is processed using the combination of hybrid VGGNet CNN model, attention module and convolutional feature encoding module. The loss function is optimized using RDA, and the matching process is done by recursive density estimation (RDE). Finally, based on the feedback obtained from the user, the RF mechanism discards the images that are non-relevant and only relevant images are retrieved. The tool used for implementation is PYTHON platform. Hence, the extensive tests on University of California Merced (UCM) and Wuhan University Remote Sensing images with 19 class (WHU-RS19) databases reveal that the proposed model performs better than several state-of-the-art techniques.

摘要

遥感（RS）领域的进步提供了大量具有更高分辨率的RS图像。如今，由于遥感影像的内容复杂，特征特殊，遥感影像检索（RSIR）已成为研究人员的一项艰巨任务。基于内容的图像检索（CBIR）方法创建了用于挖掘大型RS图像数据库的强大工具。基于内容的RSIR旨在基于从大型RS图像库发出的查询来获取具有相似视觉内容的图像。先前的大多数工作都使用预训练卷积神经网络（CNN）来形成用于RS场景分类的场景插图。在这项工作中，提出了一种有效的RSIR，该方法使用带有马鹿算法（RDA）的混合VGGNet（视觉几何组网络）CNN来基于查询图像进行RS图像的适当检索。提出的混合VGGNet CNN模型集成了降维（DR），特征提取（FE），损失函数优化，匹配过程和相关性反馈（RF）机制。用于减少尺寸的技术是主成分分析（PCA）。接下来，使用混合VGGNet CNN模型，注意力模块和卷积特征编码模块的组合来处理特征提取。使用RDA优化损耗函数，并通过递归密度估计（RDE）完成匹配过程。最后，根据从用户那里获得的反馈，RF机制会丢弃不相关的图像，仅检索相关的图像。用于实现的工具是PYTHON平台。因此，在带有19类数据库（WHU-RS19）的加利福尼亚默塞德大学（UCM）和武汉大学遥感图像上进行的广泛测试表明，该模型的性能优于几种最新技术。