Multiclass amber gemstones classification with various segmentation and committee strategies

The amber gemstones classification system is proposed and described in this paper. The amber data used in experiments are collected by amber art craft industry experts and divided manually into 30 classes. The presented investigations were care out in order to find out most accurate and fast classifier for online amber sorting application. QDA, KNN, RBF, and decision tree classifiers were tested. The descriptive features of amber were chosen as the mean, standard deviation, kurtosis, and skewness calculated on amber pixels from grayscale and HSV color spaces. The best classification result in terms of accuracy and computational performance based on the features calculation on the all pixels of sample was 60.30 % 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 best acquired result was 71.31 %. Then the final improvement was introduced by forming a committee of decision tree classifiers with Half&Half method which increased accuracy up to 73.18 %.