Cross-modal Remote Sensing Image Retrieval

Abstract

Most of the conventional remote sensing (RS) retrieval approaches used today are often based on a single modality data framework. In today�s date, the need for multimodal and cross modal based approaches especially in the RS retrieval area are growing evident with more and more data being acquired from different satellite sensors. This thesis presents a few shot learning based cross modal image retrieval framework for RS images. Few shot learning was incorporated to account for label scarcity or when the data available is insufficient and Deep CORAL loss was further integrated for domain adaptation of the cross modal data. In addition, a reciprocal points loss is also integrated for generating better discriminative features of images. We evaluate our approach on two crosssource remote sensing image datasets by training cross modally and testing uni-modally on insufficient labeled data and achieve positive results showing our framework to be helpful.