Abstract
With the advent of modern image acquisition and sharing technologies, billions of images are added to the Internet every day. This huge repository contains useful information, but it is very hard to analyze. If labeled information is available for this data, then supervised learning techniques can be used to extract useful information. Visual pattern mining approaches provide a way to discover visual structures and patterns in an image collection without the need of any supervision.
The Internet contains images of various objects, scenes, patterns, and shapes. The majority of approaches for visual pattern discovery, on the other hand, find patterns that are related to object or scene categories.Emergent pattern mining techniques provide a way to extract generic, complex and hidden structures in images.
This thesis describes research, experiments, and analysis conducted to explore various approaches to mine emergent patterns from image collections in an unsupervised way. These approaches are based on itemset mining and graph theoretic strategies. The itemset mining strategy uses frequent itemset mining and rare itemset mining techniques to discover patterns.The mining is performed on a transactional dataset which is obtained from the BoW representation of images. The graph-based approach represents visual word co-occurrences obtained from images in a co-occurrence graph.Emergent patterns form dense clusters in this graph that are extracted using normalized cuts. The patterns that are discovered using itemset mining approaches are:stripes and parallel lines;dots and checks;bright dots;single lines;intersections; and frames. The graph based approach revealed various interesting patterns, including some patterns that are related to object categories.