We also applied metafeatures to recognising Chinese handwritten characters off-line5. We built on the work of amin:chinese, who had already developed code for thinning and stroke extraction. We used a single metafeature for a stroke. The parameter space was (startx, starty, endx, endy, length, curvature). Curvature was calculated by finding the most distant point on the line from the line between the start point and the end point. There are 40 examples per sign; with 100 classes used. Ten-fold cross validation was used. Because there was only one metafeature, and knowing that there are many different types of strokes, we allowed up to 40 centroids instead of the usual 10. The results are shown in Table 5. Clearly, voting provides superior results. Furthermore, the work compares favourably with amin:c45, where again, manual preprocessing is performed. We obtain similar accuracy results to their 9.9 per cent error.
