 the purpose of this paper is to propose a classifiers design by means of a fast parallel algorithm for training and classification of samples for material recognition. The investigated classifiers are based on spectral analysis of input signals through discrete wavelet transform (DWT). The research line is accelerating the performance of the classifiers through parallel execution of training algorithm. Classifier’s training is done by fast wavelet transform on input data and extraction of characteristic coefficients. For the development of the parallel algorithm, it has been chosen «master/slave» realization: a master process that reads input data and slave worker(processes working on samples by «first(requested, first( received» principle. The chosen parallelization method is a combination of multi(threaded processing directives OpenMP and message passing interface (MPI) for communication between processes. —Classifiers design, Parallel algorithms, Parallel implementation, Performance evaluation. I. INTRODUCTION he modern technology evolution increases performance of computer systems to achieve an increasingly large number of computing elements. Parallel architectures are becoming more affordable and widespread, and computing power continues to increase. This starts the development of new methods and models of applications that can effectively take advantage of their available hardware options. With the use of computer platforms scaled multi(core processors is achieved increasing the productive power to such an extent that they can be used as models for developing new applications and for the parallelization of existing problems and algorithms. This is exemplified by the development of automatic system for non(contact materials classification, providing objective, accurate and quick information about their identity. In recent Manuscript received October 31, 2010. This work was supported by Bulgarian Ministry of Education and Science, National Science Fund Grant ДРНФ(02 A. Prof. PhD Plamenka Ivanova Borovska is with the Computer Systems Department at the Technical University of Sofia, Bulgaria, phone: +359 2 965 25 24; e(mail: pborovska@ tu(sofia.bg. B. Assist. Prof. Desislava Antonova Ivanova is with the Computer Systems Department at the Technical University of Sofia, Bulgaria, phone: +359 2 965 33 85; e(mail: d_ivanova@ tu(sofia.bg years, is actively working to implement non(contact methods for classification of materials, such efforts are focused mainly on the following areas, [1]: 1. Improvement of sensor modules, adapted for real time usage, as many of them are based on the non(contact interface with the material by electromagnetic waves; 2. Creating new and improving existing mathematical models and statistical methods and algorithms implemented them for the extraction, transformation and use of information obtained for the classification of materials; 3. Design and construction of intelligent signal classifier; 4. System testing under different conditions and processes; The paper presents results from a study of automatic classifiers ( the cores of a sorting machine for the materials recognition defining the greatest extent their physical feasibility and effectiveness. The research line is oriented to accelerating the classifiers performance through parallel execution of key elements of training algorithm. The investigated classifiers are based on spectral analysis of input signals through discrete wavelet transform (DWT), [2]. In the next session the classifiers design and implementation on the heterogeneous computer cluster will be discussed. Finally, some results analysis and performance evaluation will be done. II. CLASSIFIERS DESIGN AND IMPLEMENTATION ON A HETEROGENEOUS COMPUTER CLUSTER Automation of synthesis process of classifiers, a large amount of input data should be processed in order to extract characteristic coefficients. For that purpose each sample form given extract must be processed by wavelet transformation. There are two possible approaches: A) sequential transformation of all samples of input extract by parallel wavelet transform, using all processors of the cluster; B) parallel transformation of input data set by sequential realization of wavelet transformation, where each processor executes a separate instance of the input extract. Method A could efficient in cases, where the input set of samples is small and time for processing of each one is long. Moreover, without specialized scheme for load balancing, this approach would not be efficient if realized on heterogeneous clusters like the one used in TU(Sofia. Classifiers design and implementation for material recognition on a heterogeneous computer cluster Plamenka Borovska, Desislava Ivanova T Advances in Communications, Computers, Systems, Circuits and Devices ISBN: 978-960-474-250-9 37