Statistical Shape Descriptors for Ancient Maya Hieroglyphs Analysis

The preservation, analysis, and management of valuable and fragile historical and cultural materials with digital technologies is a field of multidisciplinary nature, and significant to the society at large. The benefits of using visual and multimedia analysis techniques in this domain are manifold. First, automatic and semi-automatic content-based analysis methods can provide scholars in the humanities (historians, anthropologists, archaeologists, and linguists) and the arts (curators, art historians, and photographers) with tools to facilitate some of their daily work, e.g., consulting, organizing, annotating, and cataloging pieces. Second, these techniques can help obtain new insights about specific theories in the archaeological field through the recognition and discovery of patterns and connections within and across pieces in a collection. Third, automated analysis techniques can boost the creation of educational systems for public access and retrieval of digital versions of ancient media. Furthermore, the careful and efficient use of these digital collections, with potential impact in local development and tourism, also has a definite economic value. This dissertation presents an interdisciplinary approach between computer vision and archaeology towards automatic visual analysis of ancient Maya media, more specifically of hieroglyphs. The ancient Maya civilization has been regarded as one of the major cultural developments that took place in the New World, as reflected by their impressive achievements, encompassing the artistic, architectural, astronomical, and agricultural realms. Paramount among these is their refinement of a fully-phonetic writing system that ranks among the most visually sophisticated ever created in world history. Therefore, our work is guided by realistic needs of archaeologists and scholars who critically need support for search and retrieval tasks in large Maya imagery collections. More precisely, we address the problems of statistical shape description and Content-Based Image Retrieval of Maya hieroglyphs. The type of data we analyze is rich in visual information and exhibits high degrees of visual complexity. Furthermore, the elements of the ancient Maya writing system often present inter-class visual similarity. In this dissertation, we first present an evaluation of state-of-the-art shape descriptors for the task of shape-based image retrieval. We then introduce the Histogram-of-Orientations Shape-Context, which is a new shape descriptor designed to overcome certain drawbacks found in state-of-the-art methods, and we demonstrate its potential to deal with shapes originated from different data sources. Moreover, we present the results of using shape descriptors towards the statistical analysis of the visual evolution of Maya hieroglyphs over time and across regions of the ancient Maya world, as well as for the statistical analysis of the intra-class and inter-class visual variability of syllabic classes of Maya hieroglyphs. We also compare the performance of clustering and sparse coding techniques in the construction of efficient representations of shapes. Finally, we present results on detection of segmented symbols in large inscription based on shape descriptions.