Automatic amber gemstones identification by color and shape visual properties

• Published in: Engineering applications of artificial intelligence 2015-01, Vol.37, p.258-267
• Main Authors: Sinkevicius, Saulius, Lipnickas, Arunas, Rimkus, Kestas
• Format: Article
• Language: English
• Subjects: Centroids; Classification; Classifiers; Color; Decision trees; Expert systems; Gems; Image classification; Image matching; Mathematical analysis; Pixels; Supervised learning
• ISSN: 0952-1976; 1873-6769
• DOI: 10.1016/j.engappai.2014.09.011

This paper proposes and describes novel techniques for the amber gemstones labeling system. The amber data used in experiments are collected by amber art craft industry experts and the presented investigations were carried out in order to develop a classifier for online amber sorting application. Amber pieces are identified and labeled to one of 30 color classes or to one of 20 geometric shape classes. For identification Quadratic Discriminant Analysis, K Nearest Neighbors, Radial Basis Function, Naive Bayes, Decision Tree, and pruned Decision Tree classifiers were tested. As color descriptive features mean, standard deviation, kurtosis, and skewness calculated on amber pixels from grayscale and HSV color spaces were chosen. The best classification result with the features calculation on all the pixels of sample was 69.29% accuracy, obtained by Pruned Decision Tree classifier. In order to improve the classification results, the pixels of amber samples were grouped into predefined concentric ring segments and the accuracy rose by 10%. Then the final improvement was introduced by forming a committee of Decision Tree classifiers with Half&Half method which increased accuracy up to 81.60%. For shapes identification the Centroid Distance Function was selected as it preserves the order of landmark points. Using labeled samples the Decision Tree classifier was trained. The training of classifier was made by acquiring all possible orientations of Centroid Distance Function for each image in training set and then feeding them to Decision Tree. In the classification step all the shifted and flipped Centroid Distance Function variations of the testing sample are voting for the class using the trained Decision Tree. Experimental results have shown that the proposed technique is effective in organic shapes classification to selected geometric shapes even if there is high ambiguity between organic shapes and 72.10% accuracy was acquired. Both proposed classifiers can be used in real time application independently or in combination.