Computing Invariants for Structural Change Detection in Urban Areas

This paper introduces a new approach to compute similarity/dissimilarity between images or regions from images. It borrows concepts from object recognition and content based image retrieval. The main idea is to build a bag-of-features where the feature clusters are learnt from a single pair of images. Normalised cut is used to segment the image into regions which exhibit homogeneous and regular structural pattern distribution. Histograms of feature types are compared using the chi² distance. The application is on very high resolution optical satellite images, for urban development analysis.