Cultural Heritage Applications: Archaeological Ceramics and Building Restoration - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 141

this period, the productive strategy of household chose intentionally to simplify the
production process. Simple ways of ceramic manufacture and cooking were
employed to obtain kitchen's recipients with thermal shock resistance. Thus,
ceramics were manually made from little-decanted clays and cooked at low
temperatures. The results were coarse pieces from the Middle Ages with physical
properties comparable to the Bronze Age pieces [ 24 ].
Let us analyze the ultrasound-based results from the point of view of the
porosity and density. We observed that the porosity and density of the Bronze Age
pieces are relatively close to porosity and density of the pieces from Roman and
Middle Ages. This explains why 7 and 14 % of the Bronze Age pieces were
assigned to the Roman and Middle Ages periods in Table 6.3 . Similarly, the pieces
from the Iberian period and the Middle Ages have similar porosities and densities,
so this may justify why 2 % of the Iberian pieces were assigned to the Middle
Ages. The 9 % of pieces of the Iberian period that should have been assigned to
the Roman period were incorrectly assigned because the Iberian ceramic is very
close to one of the three kinds of Roman ceramics (sigillata, common, and
amphora)—the common kind—. This also explains why the corresponding 19 %
of pieces of the Roman period were incorrectly assigned to the Iberian period. No
clear explanation exists for the lack of symmetry in the confusion matrix of
Table 6.3 ; however, it must be taken into account that the training process
introduced some degree of arbitrariness because of the probabilistic labelling of
the expert. Thus, it seems that the expert was able to clearly identify the pieces
from Iberian and Middle Ages, but had more difficulties with the Bronze Age and
Roman ones. This uncertainty may have been transmitted to the classifier during
the training stage.
The experiments with standard methods of ceramic characterization used in
archaeology not only show that correlations between the extracted parameters
from the ultrasound signals and the physical properties of the materials were
found. Moreover, they also have demonstrated some advantages of the proposed
ultrasound method. The equipment required for NDT by ultrasound is, in general,
less costly, and the experiments are easier to perform. The pieces are not damaged
in any way during testing, nor is it necessary to alter or destroy any of the material
that is analyzed. Very significant differences for the time required to analyze the
pieces were demonstrated: the ultrasound analysis (measuring, processing, and
automatic classification) for 480 pieces took only 6 h; the SEM analysis (tube
preparation and electron microscope analysis) for 80 pieces took 274 h; the
porosity and density analyses (immersion and weighing of the pieces) for 80 pieces
took 288 h.
There are limitations to the application of this procedure due to the fact that the
training of the classifier must be performed from a specific set of data. Thus, the
classifier must be adapted to a specific data model and its efficiency is restricted by
the fact that the new data to be classified must follow the same data model.
Nevertheless, the training of the classifier could progressively be improved by
increasing the number of pieces for each known chronological period. With proper
training, the classifier would be able to provide a prediction of the chronological