Content Based Image Retrieval on Satellite Imagery

In this study, a content-based image retrieval system for satellite imagery is proposed. The main purpose of the paper is to find the most relevant images to the query image from a massive satellite imagery database using deep learning methodology. For this purpose, graph convolutional network and hash network are combined for image retrieval tasks. Graph convolutional network is used for graph-based image representation. The graph-based image representations are fed into multi-layer perceptron to learn hash code of the image. In this way, hash code based final image representations are obtained. The image searching is based on nearest neighbor approach with the use of these representations. The proposed model is applied to UC Merced and BigEarthNet datasets with extensive experiments.