IMAGING ROMAN STILUS TABLETS
A.K.Bowman and J.M.Brady, University of Oxford
In addition to an exceptional quantity of well-preserved ink writing-tablets, Vindolanda has also produced a large number of stilus tablets; the 46 examples discovered in the 1990s bring the total to around 200, of which many have visible remains of writing. There are, of course, a few which are relatively straightforward to read (particularly the addresses on the back), but the majority of tablets from Vindolanda and in museums and collections elsewhere remain intractable and unread. The problems posed by the majority of such stilus tablets, in which the wax surface has perished and we are left with traces of incision where the stilus penetrated the wax to the wood beneath, are familiar. We cannot read them because we cannot see the incisions clearly enough; the incisions are often incomplete especially at the ends of strokes; they are often palimpsest; the visibility of the text is undermined by wood grain and other casual damage.
In 1997 a project was initiated at the University of Oxford by the Centre for the Study of Ancient Documents in conjunction with the Department of Engineering Science to develop a new computer-based image-enhancement technique for incised material. The project directors are Dr.A.K.Bowman and Professor J.M.Brady and the work is funded for 1998-2001 by the UK Engineering and Physical Sciences Research Council.
The classicists and archaeologists involved in the project are:
The engineering scientists (all of the Department of Engineering Science) are:
For preliminary accounts see A.K.Bowman, J.M.Brady, R.S.O.Tomlin “Imaging Incised Documents,” Literary and Linguistic Computing Vol.12, no.3 (M.Deegan, A.K.Bowman (ed.), Imaging Documents, Oxford 1997, 169-76); a summary description of the project can also be found here. An account of the engineering issues will shortly appear as N.Molton, Xiaobo Pan, J.M.Brady etc., “Visual enhancement of inscribed text,” (forthcoming)
The main factors which make the writing on stilus tablets difficult to identify and transcribe are: (i) the scratches are shallow and they can often only be seen properly under intense low-raking light; (ii) the background on which the scratches are made almost always has a comparatively heavy wood grain (emphasised by the fact that the softer parts of the wood have often degraded more than the harder ones, producing an undulating surface) which constitutes a distracting background texture; (iii) incidental and casual pitting further complicate the task of isolating the writing ; and (iv) the stilus tablets are frequently palimpsest, thus offering two or more overlaid texts which need to be distinguished.
This unique example from Vindolanda was excavated with the wax coating intact. The conservation process dissolved the wax, affording opportunity to compare the text on the wax with the scratches left on the wood below.
In our attempts to produce more readable texts we have addressed two main issues:
1. Woodgrain removal
The woodgrain is usually more or less uniform over the surface of a single stilus tablet and it is a relatively large visible feature compared to the incisions we want to read. By aligning the camera appropriately when capturing digital images of the tablet we can orientate the image so that the woodgrain is aligned with the horizontal. This particular property is then exploited in the 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. Currently the most efficient way to do this is to mask out the values along the vertical of the (local) Fourier transform of the image. The resulting images contain much less distracting information and thus produce a great improvement for the human reader. During the early stages of the project we have succeeded in producing images in which the distraction of the wood-grain signals has been very significantly reduced.
2. Finding incisions
The incisions which are of interest to us 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, with only a small range of grey values, which makes them very difficult to distinguish from background ‘noise’. Since the size of the incisions is very small compared to the distance between the stilus tablet and the indirect measuring tool (such as a stereo-camera) and inherently unsuitable for direct measuring methods (such as confocal scanning optical microscopy), another class of indirect ranging techniques is used (Shape from Shading), in which three-dimensional surface shape is computed from variations in shading.
The technique which we have so far developed in this project combines the key properties of photometric stereo and shadows. When someone carries a torch on a dark night slight movements of the torch in the hand will intermittently induce rapid movements of cast shadows, so that the object casting the shadow seems to jump out from the background. The extent to which it does so depends on the orientation of the light beam relative to the orientation of the surface on which it casts the shadow and its separation in depth from the surrounding background surface. A simple qualitative technique (movement or not) suffices to discriminate between incisions and surface discolourations, but more detailed information is needed to discriminate between incisions from two distinct sources (such as two hands on a palimpsest tablet).
We have called this technique Phase Congruency Shadow Stereo and have developed the following procedure. As a broad approximation to the ‘manual’ approach traditionally used by the palaeographer we capture multiple images of the same stilus tablet with a light source moving in an arc ‘over’ the tablet, beginning at a very low-raking angle (about 10 degrees above surface) and moving up in 5 degree steps. Correctly positioned lighting provides clear shadows, reduces the effect of noise and generates the maximum amount of information about the incision structure. The imaging arrangement is designed so that the edges of the incisions proximal to the light-source cast shadows, while those distal to the light-source show a transition from a shadow to a highlight. The crucial image-processing task in the application is to identify the point at which the key features (incisions) make the transition from shadow to highlight; when this has been done for each individual image, as far as possible, the output from all processed images is then combined in order to detect: (i) the movement of certain features corresponding to a moving shadow-highlight combination which moves as the light-source is moved up and (ii) the fact that certain features do not move although the light-source is moving. The latter can then be identified as two-dimensional features such as discolourations) rather than the three-dimensional features (incised letters) which we are attempting to isolate. It is particularly important to exploit the fact that surface discolourations, which can easily be confused with shadows in a single image, do not move and we have so far used a simple correlation technique to discriminate between surface discolourations (stationary features) and incisions (moving features). The results are very encouraging, giving a good classification and greatly reducing spurious responses due to ‘noise’.
This, in our view, represents a considerable advance in technique compared to the standard edge-detector software which does not discriminate sufficiently between deliberate incision and casual damage to the surface of the tablet. In our future research we propose to develop this aspect further, incorporating both the crucial Phase Congruency characteristics of the intensity transition and recent developments in correlation-based stereo. This aspect is even more important in the next stage which involves extending the basic technique of stereoscopic imaging of shadows by combining results from multiple azimuthal angles.
Further information from:
Dr. A.K.Bowman (firstname.lastname@example.org)
Dr. C.V.Crowther (email@example.com)
Professor Michael Brady (firstname.lastname@example.org)
Last updated on Friday, 15 December, 2000: 15:44:53