CSAD Computer Vision Seminar8 November, 1997 |
A day-long seminar was held at the Centre on Saturday, 8 November to explain the aims and results of the Centre's recent work with particular emphasis on projects involving the interaction of documentary studies and developments in the field of computer vision. The seminar had both a symbolic and a practical purpose: to offer thanks to individuals and representatives of institutions which have supported our work in the past two years, to explain the progress we have made and to invite comment and advice on future progress and plans. Those who attended by invitation were: Prof. Roger Cashmore (Oxford: Chairman, University Research and Equipment Committee), Dr. David Clark, (Director, Engineering and Science, EPSRC), Sir Anthony Cleaver (Chairman, AEA Technology plc), Dr. David Cooper (Libraries Automation Services, Oxford), Professor Michael Crawford (University College London), Dr. Marilyn Deegan (Queen Elizabeth House, Oxford), Professor John Krebs (Chief Executive, Natural Environment Research Council), Professor Paul Langford, (Oxford: Chairman Designate, Humanities Research Board), Professor John Laver (Chairman, Humanities Research Board), Mr. Hamish Orr-Ewing (former Chairman of Rank Xerox UK and of Jaguar plc), Mr. Chris Peters (Oxford University Development Office), Dr. Tim Potter (Keeper of Prehistoric and Romano-British Antiquities, The British Museum), Dr. Seamus Ross (Director, Humanities Computing and Information Management, University of Glasgow).
The Centre and its projects were represented by Alan Bowman, Charles Crowther, Roger Tomlin, Dirk Obbink and Gideon Nisbet and, from the Department of Engineering Science, Professor Mike Brady, Dr. Andrew Zisserman and Veit Schenk.
The day's programme was framed by presentations of the particular problems facing documentary historians in dealing with damaged and difficult documents written on a variety of surfaces: wood, stone and papyrus. Advances in computer vision applications have been deployed to resolve analogous problems in other fields-in medicine, for example, or structural engineering. The the focus of the seminar was on how similar techniques might be applied to the decipherment of intractable ancient texts. Epigraphers and papyrologists have developed their own practical techniques for working with problematic texts and the integration of this body of expert knowledge into the design of image processing applications can offer insights that may lead, in turn, to progress in other applications of computer vision.
The central section of the seminar was an exposition by Mike Brady and Veit Schenk of how this interaction works in practice in the Centre's current project on wooden stilus tablets from Roman Britain:
"The aim of our work is to combine the historian's knowledge, skills and techniques with modern computing technology in order to develop tools which facilitate the reading of some tablets and extend the range of documents that would otherwise remain illegible.
A variety of factors makes stilus tablets particularly difficult to interpret: the letter traces are shallow and they can only be seen properly under intense low-raking light; the background on which the incisions were made almost always has a comparatively heavy wood-grain producing a distracting background texture-emphasised by the fact that the softer parts of the wood tend to degrade more than the harder ones resulting in a wave-like undulating surface; incidental and casual pitting further complicate the enhancement of writing; and, as if all of this was not already bad enough, a single tablet would often have been used more than once, thus offering two or more overlaid texts that need to be distinguished.
So, in order to help the historians read stilus tablets more effectively there are several issues we have had to address: we want to be able to remove woodgrain and other distracting features so that the end-user can work on 'clean' images containing 'useful' information only. We also need to be able to isolate features that correspond to incisions and give the end-user the tools to interactively accept or reject and manipulate these features on a computer in the process of decipherment.
Woodgrain removal: The woodgrain on most stilus tablets is usually more or less uniform over the surface of the tablet and is relatively large compared to the incisions we are looking for. When taking images we arrange for the woodgrain to be aligned in the horizontal direction. This particular property is then exploited in our process of woodgrain removal which is based on masking out low-frequency components in the vertical direction, keeping all high-frequency information corresponding to incisions. The resulting images contain much less distracting information and are a great improvement for the human reader.
Finding incisions: The incisions the historian is interested in typically measure 0.5 mm across and are of varying but shallow depth, to a maximum of 1mm. Visually, the incisions are of low contrast, often just a few grey values, so that they are very difficult to distinguish from background noise. We employ a technique which models the working practice of the epigrapher by combining the key properties of photometric stereo and shadows. In practice, the scholar 'reading' a stilus tablet uses low-angle raking light: the tablet is moved 'over' and 'around' the light source, casting 'moving' shadows where the incisions are approximately perpendicular to the incoming light. The 'over'-movement is used to determine whether there are incisions in this particular direction, i.e. to distinguish incisions from other surface markings. The 'around'-movement is used to find incisions in varying orientations on the tablet. To model this 'manual' technique we take a number of pictures of the same tablet with a light source moving in an arc 'over' it, beginning at a very low-raking angle and moving up in 5 degree steps. A major advantage of the computerised over the manual approach is that the historian no longer has to remember where exactly an incision is located before moving the light-source or to draw it on paper in order to combine individual scratches into a letter or coherent writing; the results of the 'around'-movement are 'integrated' in the computer.
The major problem to overcome is to find the appropriate features in each image: current techniques such as histogram equalisation, standard edge finders, and anisotropic diffusion give poor results because of the low-contrast and high-noise content of the images. We have had promising results, however, using image features that correspond to locations of high phase congruency (PC). Current implementations of PC are determined by a range of specific design choices. We believe that a theoretical and practical analysis within the context of our stilus tablets application will lead to greatly improved performance of PC both in this and in other applications."
The Centre's work on wooden stilus tablets is breaking new ground in
developing techniques of imaging damaged surfaces which will eventually
be applicable to many different kinds of material-from deciphering other
types of ancient document, to spotting hairline flaws in turbine blades
or examining the ducts and structures in the human heart.
| Created on |