VOL. 10, NO 19, OCTOBER, 2015 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences ©2006-2015 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com 8853 PSO-BP ALGORITHM IMPLEMENTATION FOR MATERIAL SURFACE IMAGE IDENTIFICATION Fathin Liyana Zainudin 1 , Abd Kadir Mahamad 1 , Sharifah Saon 1 , and Musli Nizam Yahya 2 1 Faculty of Electrical and Electronics Engineering, Universiti Tun Hussein Onn Malaysia, Parit Raja, Johor, Malaysia 2 Faculty of Mechanical and Manufacturing Engineering, Universiti Tun Hussein Onn Malaysia, Parit Raja, Johor, Malaysia E-Mail: fathin.liyana.zainudin@gmail.com ABSTRACT Implementation of neural network for acoustic computation is not new. In this paper, a new improved method in predicting material surface from photographic image was implemented using a hybrid of particle swarm optimization and back-propagation neural network (PSO-BP) algorithm. Before the system classified the data using PSO-BP algorithm, the photographic images of room surfaces need to be extracted using Gray Level Co-occurrence Matrix (GLCM) and Modified Zernike Moments. The result indicated that the PSO-BP algorithm have a higher accuracy compared to the BP algorithm, managed to record highest accuracy of 88% as opposed to 81.3% for the latter. Keywords: particle swarm optimization, back-propagation, image processing. INTRODUCTION Material type is an important feature in room acoustic engineering; from determining the absorption coefficient of said material to computation of the room reverberation time. From the photographic images, the texture of the surfaces whether ripple, rough, smooth, etc. are captured. In analyzing the texture, the first and most important task is to extract texture features which have all the information about the textural characteristics of the original image. Previously, a few researches were performed using various types of image processing and image classification methods in building the system (Zainudin et al. 2014), (Mahamad et al. 2014), (Sari, Hazli and Shimamura, 2013). In this paper, a hybrid of particle swarm optimization and back-propagation (PSO-BP) algorithm is proposed to improve and upgrade the material surface identification system. Application of feed forward neural network is actually a common practice for classification of the non- linearity separable patterns of the texture. Currently, there are many algorithms for feed forward neural network (FFNN) training for example the back-propagation (BP) algorithm, the Levenberg-Marquardt (LM) algorithm, the genetic algorithm (GA), and particle swarm optimization (PSO). Out of these algorithms, one of the most popular and commonly used is the BP algorithm. As it is actually a gradient based heuristic method where the concept of this algorithm is basically to search and move along the gradient towards the most minimum hence making the algorithm simple and easy to apply. The BP algorithm nevertheless had the disadvantages of slow convergence and easily getting stuck in the local minimum (Zhang et al. 2007) especially for non-linearity separable classification problems. Therefore, to overcome this particular problem, the PSO-BP algorithm is introduced to the system. PSO algorithm itself is proven for having a fast convergence during training although it has the drawback of easily getting stuck in the global minimum (Singh and Singh, 2012). PSO-BP algorithm basically utilize PSO algorithm to find the global optimum and BP algorithm to search for the optimal weights in order to avoid getting trap in the local minimum. PSO-BP algorithm also has the upper hand of having a better convergence speed and accuracy (Han, Gu and Ju, 2011), (Liu and Qiu 2009). PSO-BP algorithm overview a) BP algorithm BP architecture is not any different than any feed forward neural network where it is consists of 3 different layers; input, hidden, and output layer. The most significant difference is the existence of the back- propagation error that was feed each time to update the weights. BP algorithm neural network is still considered a supervised learning method where the desired output is needed to calculate the error that used for the weights update. The weights of BP algorithm will be updated with each iteration time. BP algorithm computes the squared error of the neural network, for gradient E as in equation (1). =ܧ ଵ  ∑ | − |ݕ ଶ (1) with N = number of training data, t = desired output and y = actual output. The actual value of the previous expression depends on the weights of the network. BP updates the weights by shifting them along the gradient descendent