Chercheur

Teachers Login Form

 

Admin

Pupils Login Form

 



Unsupervised classi cation of weld defects in radiographicimages based on mutivariate generalized Gaussian mixture

Auteurs : Nafaa Nacereddine, Aicha Baya Goumeidane, Djemel Ziou
Année : 2016
Domaine : Electronique
Type : Communication
Conférence: The International Conference on “Imaging, Vision and Learning based on Optimization and PDEs”, IVLOPDE2016
Résumé en PDF : (résumé en pdf)
Fulltext en PDF :
Mots clés : Mixture model, Multivariate GGD, radiography, weld defect, classification

Résumé :

An accurate modeling of unknown probability density functions pdfs ofdata, encountered in practical applications, can play an important role in machinelearning, clustering and pattern recognition. Including Gaussian, Laplacian anduniform distributions as special cases, multivariate generalized Gaussian distribution (MGGD) is potentially interesting for modeling the statistical properties ofcomputer vision applications. In fact, the GGD is an elliptically contoured distribution characterized not only by a mean vector and a scatter matrix , but also bya shape parameter determining the peakiness of the distribution and the heaviness of its tails making this distribution more flexible than multivariate Gaussiandistribution (MGD) and thus, more suitable for modeling, among others, imagesor features extracted from these images. However, the expressions of the partialderivative equations (PDE) deriving the MGGD parameters handle highly nonlinear functions including piece-wise, logarithm, gamma, psi, power, etc. So, a particular attention is required for the derivatives computation, especially, for the matrixdifferentiation. Here, the solutions are given by the Newton-Raphson method. Inorder to carry out the experiments, hundreds of weld defect regions, extractedfrom weld radiographic films provided by the International Institute of Welding,are used. These defects represent four weld defect classes (crack, lack of penetration, porosity, solid inclusion) and are indexed by eight geometric measures. Theexperimental results, in terms of confusion matrix and total classification rate,demonstrate an outstanding performance of the MGGD-based mixture model inthe weld defect data modeling compared to the multivariate Gaussian mixturemodel.