Using supervised machine learning to classify ceramic fabrics

Abstract

Objective: The purpose of this thesis was to assess the ability of supervised machine learning to discriminate between images of ceramic fabrics and evaluate the requirements for creating an effective dataset.

Method: Weighted Neighbour Distance using Compound Hierarchy of Algorithms Rep- resenting Morphology (wndchrm) algorithm, was applied to Zulu ceramic fabrics from South Africa. Wndchrm was used to extract thousands of image content descriptors, assign weights to the extracted features by learning their discriminative power from training examples, and classify unlabelled images by searching for a class with the nearest distance to the mean of the feature vector.

Results: Wndchrm was successful in distinguishing ceramic fabrics by differences in paste. Comparable results were obtained in separate experiments designed to identify differences in paste by region, community and individual potters. In all cases, a sample size of 50 train- ing images per class was sufficient to produce 90% to 95% accuracy. In contrast, the experi- ments meant to identify fabrics by shaping techniques, represented by coiling and slab building, reached the accuracy of only 65% to 70%.

Conclusion: The experiments show that wndchrm is an effective method for classification of ceramic fabrics, which can increase consistency in comparing and classifying ceramic fabrics by variation in paste. Creating and sharing training dataset libraries is the next step necessary for wide adoption of supervised machine learning in this field.