An Integrated Region and Patch Feature Descriptor Model for Efficient Retrieval of Remote Sensing Images

Abstract

The region-based descriptors face difficulties representing the complex thematic classification of remote sensing images (RSI) with heterogeneous objects and backgrounds.  The retrieval task using large archives thus becomes cumbersome on these images.  Also, the region-based classification methods cannot consistently identify interest points on their boundaries.  This work recommends a new object-based image retrieval framework incorporating a region and patch feature descriptor (RPFD) for high-resolution remote sensing image retrieval (HR-RSIR) tasks.  The method effectively combines local patch and region-based features into a region-based framework.  The regional context features (RCF) are captured through an efficient seeded region growing segmentation (SRGSeg).  The RCF integrates the augmented features from patches near the regions for better classification.  The proposed integrated region and patch feature descriptor framework for RSIR (IRPFD-RSIR) focuses on reliable texture modeling at the region level, with an augmented patch-based feature descriptor to get more understanding of complex RSI.  A weighted k-nearest neighbor (Wk-NN) classifier is incorporated for searching similar classes, utilizing an effective feature selection method.  Three benchmark RSI datasets are used for evaluating the framework.  The results compared with various feature descriptors and frameworks substantiate the overall efficiency of our retrieval model in terms of precision, recall, and F1-score.