A hierarchical image processing approach to analysis of early medieval manuscript art

Abstract

The discipline of art history is one in which digital technologies have traditionally played a minor role. Recently with advances in computing methods, software and the capabilities of hardware, art analysis has begun to see the benefits of automated image understanding and pattern recognition techniques. The specific domain of historical manuscript analysis aims to answer questions regarding the visual design, structure, execution, comparison, meaning and context of these historical works of art. Compared to other forms of art, such as paintings, architecture, and abstract art, investigations into manuscripts deal with the construction of design at the much more primitive level of textual and illustration content and styles. Such a bottom-up approach to analysis is encouraging for digital image processing studies into art. The range of questions that are of interest to art historians extends from the elicitation of design and layout, the accuracy of execution and the restoration of degraded art works to the understanding of the aesthetic effect of these manuscripts. Recent work in the field of digital art history has seen application of a various pattern recognition and image statistical techniques to specific art historical analyses. Most of these applications are based on techniques that are irreconcilable with each other methodologically. The subject of this thesis is the problem of developing a unified framework for such studies that allows the historian to answer a wide range of art historical problems. The rudiments of such a framework are developed herein and its applicability is demonstrated via approaching some questions regarding the Book of Kells (as an exemplar), specifically the analysis of calligraphy, interlace, detection of various design elements and degradation removal. The framework improves on biologically inspired hierarchical models of vision and makes it feasible to develop art historical applications in a consistent fashion. This consistency also affords the possibility of approaching problems of a highly abstract nature, viz. aesthetics, using information extracted from other (less abstract) applications, such as object detection. As a particular application, the identification of areas of manuscripts that might draw more attention than others is demonstrated.