[2]P. Clemente-Perez, V. Garceran-Hernandez, H. D. Puyosa-Pina, and L. M. Tomas-Balibrea, "Automatic system to quality control: Using artificial vision and neural nets for classification of marble slabs in production line", Proc. Int. Symp. Artificial Neural Networks, pp. 1995
[3]M. Deviren, M. K. Balci, U. M. Leloglu, and M. Severcan, "A feature extraction method for marble tile classification", Proc. 3rd Int. Conf. Computer Vision, Pattern Recognition, and Image Processing, pp. 2000 .
[4]J. Chang, G. Han, J. M. Valverde, N. C. Griswold, J. F. Duque-Carrillo, and E. Sanchez-Sinencio, "Cork quality classification system using a unified image processing and fuzzy-neural network methodology", IEEE Trans. Neural Netw., vol. 8, no. 4, pp. 1997
[5]P. R. Drake and M. S. Packianather, "A decision tree of neural networks for classifying images of wood veneer", Int. J. Adv. Manuf. Technol., vol. 14, pp. 1998.
[6]M. Unser, "Sum and difference histograms for texture classification", IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-8, no. 1, pp. 1986 .
The system consists of a mechatronic prototype(physical components for acquiring the marble slabs images in suitable lighting conditions) and computational algorithms(analysing the colour, texture of the surface and classifying them into their corresponding groups). Four colour spaces namely RGB, XYZ, YIQ and K-L are taken into consideration. Each colour space is a three dimensional gray level. Texture analysis is done based on the gray level dependence between adjacent pixel using the sum and difference histogram algorithm. Seven statistical features such as mean, variance, energy, correlation, entropy, contrast and homogeneity are then calculated from these histograms. In order to reduce the 21parameters(number of dimension*number of statistical feature=3*7) a feature extraction process is implemented using the principal component analysis. This gives a much smaller data(f) than the original without redundant information. The number of features is chosen such that their individual principal components contributes to over 0.5% of the total variance of the data set. Finally a neural network multilayer perceptron with backpropogation and adaptive learning rate is used to classify the marble slabs. The structure of the neural network consists of a three feedforward layers 1. The input layer with f neurons 2.the hidden layer 3.the output layer with 3 neurons (one for each category). The activation functions chosen are the hyperbolic tangent for the hidden layer and the linear function for the output layer. The parameters used where number of iterations=3000epochs, 1 cross validation, goal error=0.0001, maximum performance increase threshold=1.04, learning rate increasing ratio=1.05 and decreasing ratio=0.7.
The performance of the classifier was evaluated based on the percentage of correct classification, false positive and false negatives. The results showed that slabs of low quality were correctly classified than slabs of extra/commercial quality. Further RGB and XYZ had similar patter distribution and they both achieved poor results (only three quarters of the sample were correctly classified). In contrast, YIQ and KL achieved good results with high rates of correct classified patterns(above 98%). This is because YIQ and K-L provide texture and colour information in different channels, so the texture information needed to classify marble images according to their quality is extracted better than in the case of RGB and XYZ which do not provide such a clear separation between texture and colour. This automatic classification system increases the homogeneity of marbles supplied and reduces time, cost and discrepancies in the marble industry. This system can also be useful for other natural product like granites and wood.