Pattern extraction from data with application to image processing

Abstract

The term Information Extraction refers to the automatic extraction of structured information from data. In such a context, the task of pattern extraction plays a key role, as it allows to identify particular trends and recurring structures of interest to a given user. For this reason, pattern extraction techniques are available in a wide range of applications, such as enterprise applications, personal information management, web oriented and scientific applications. In this thesis, analysis is focused on pattern extraction techniques from images and from political data. Patterns in image processing are defined as features derived from the subdivision of the image in regions or objects and several techniques have been introduced in the literature for extracting these kinds of features. Specifically, image segmentation approaches divide an image in ”uniform” region patterns and both boundary detection and region-clustering based algorithms have been adopted to solve this problem. A drawback of these methods is that the number of clusters must be predetermined. Furthermore, evolutionary techniques have been successfully applied to the problem of image segmentation. However, one of the main problems of such approaches is the determination of the number of regions, that cannot be changed during execution. Consequently, we formalize a new genetic graph-based image segmentation algorithm that, thanks to the new fitness function, a new concept of neighborhood of pixels and the genetic representation, is able to partition images without the need to set a priori the number of segments. On the other hand, some image compression algorithms, recently proposed in literature, extract image patterns for performing compression, such as extensions to 2D of the classical Lempel-Ziv parses, where repeated occurrences of a pattern are substituted by a pointer to that pattern. However, they require a preliminary linearization of the image and a consequent extraction of linear patterns. This could miss some 2D recurrent structures which are present inside the image. We propose here a new technique of image compression which extracts 2D motif patterns from the image in which also some pixels are omitted in order to increase the gain in compression and which uses these patterns to perform compression. About pattern extraction in political science, it consists in detecting voter profiles, ideological positions and political interactions from political data. Some proposed pattern extraction techniques analyze the Finnish Parliament and the United States Senate in order to discover political trends. Specifically, hierarchical clustering has been employed to discover meaningful groups of senators inside the United States Senate. Furthermore, different methods of community detection, based on the concept of modularity, have been used to detect the hierarchical and modular design of the networks of U.S. parliamentarians. In addition, SVD has been applied to analyze the votes of the U.S. House of Representatives. In this thesis, we analyze the Italian Parliament by using different tools coming from Data Mining and Network Analysis with the aim of characterizing the changes occurred inside the Parliament, without any prior knowledge about the ideology or political affiliation of its representatives, but considering only the votes cast by each parliamentarian.