TOWARDS A COMPUTER VISION BASED QUALITY ASSESSMENT OF TAHITIAN PEARLS: Automatic Nacre Thickness Measurement and Color Classification
Year: 2016
Type: dissertation
Abstract: The nacre thickness measurement of Tahitian pearls is part of an obligatory quality control for pearls deemed for exportation. It is currently done manually by government experts for over 7 million pearls annually with an export volume of over 61 million €. In Part I of this thesis our algorithm to automatize this procedure under consideration of the unique features of Tahitian pearls is presented. The developed algorithm was tested on X-ray images of 298 Tahitian pearls that were classified manually by government experts. The detection accuracy of our algorithm reaches 98% in its basic form and 100% with a proposed optimized version. A prototype was developed and is currently implemented at the governmental institution in charge of the obligatory control. In Part II of this thesis, our work on normalized rgb color space theory and the classification of Tahitian pearls based on their perceived color is presented. A quintessence of the presented work is the formalization of chromatic index calculation in the context of normalized rgb histogram binning. It is shown that every chromatic index based on a linear combination of normalized rgb channels can be expressed by a single variable; an angle that corresponds to the human perceptual parameter Hue. Experimental classifications based on Artificial Neural Networks and chromatic indices over the whole possible range of the normalized rgb color space were conducted by classifying 150 Tahitian pearls that were labeled by 7 experts. The results show that the choice of chromatic indices affects significantly the classification performance of an ANN with a given topology. The global performance varies within a range of up to 20% over the tested feature vectors. Furthermore does the classification accuracy of each class depend on the chosen index, with variations up to 100% between different feature vectors. These findings are of great importance for computer vision applied to color object classification, as currently only a handful of chromatic indices are used in the literature, which is a drastic limitation of possible classification results. Additionally, a new binning strategy is proposed that takes into account the topology of the normalized rgb color space. The application of the proposed topological binning has in tendency a positive effect on the global classification rate when compared to standard equidistant binning. For certain human classifications, the accuracy difference is significant with variations of up to 18%. Furthermore, it is shown that it is generally feasible to classify Tahitian pearls based on their perceived color. Classification rates of over 84% for training and over 79% for test data are reached for all 8 used human classifications.
Source: HAL
Author: Loesdau, Martin
Institution: University of French Polynesia
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2656793834