International Review on Computers and Software (I.RE.CO.S.), Vol. 9, N. 6 ISSN 1828-6003 June 2014 Manuscript received and revised May 2014, accepted June 2014 Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved 1007 Artificial Neural Network-Based Texture Classification Using Reduced Multidirectional Gabor Features Mohammed W. Ashour, Fatimah Khalid, Lili N. Abdullah, Alfian A. Halin Abstract – In this paper, a technique to classify Engineering Machined Textures (EMT) into the six classes of Turning, Grinding, Horizontal-Milling, Vertical-Milling, Lapping and Shaping, is presented. Multidirectional Gabor features are firstly extracted from each image followed by a dimensionality reduction step using Principal Components Analysis (PCA). The images are finally classified using a supervised Artificial Neural Network (ANN) classifier. Experimental results using a 72-image dataset demonstrate that PCA is able to reduce computational time while improving classification accuracy. In addition, the use of the proposed Gabor filter shows to be more robust compared to other existing techniques. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Texture Features Extraction, Features Reduction, ANN Classification, Gabor Filter, Principal Component Analysis Nomenclature Gabor kernel Width of the Gaussian Gabor kernel G(u,v) The Gabor transform I(x, y) An image fragment Orientation of the Gabor filter A mathematical constant a ratio (3.14) h(i) Intensity-level histogram of image (i) p(i,j) The (i,j) th entry in an image matrix G Number of intensity levels in an image µ x Mean of the row sums of a matrix x Standard deviations of the row sums µ y Mean of the column sums of a matrix y Standard deviations of the column sums p Number of principal axes T 1 , T 2 , ..., T p Eigenvectors of the covariance matrix m Number of all principal components S The covariance matrix i The largest eigenvalue of S ˆ S The global covariance matrix ˆ The i th largest eigenvalue of ˆ S i x An input image from the EMT dataset c n A single class in the EMT dataset G 1 , G 2, ..., G 6 The filtered images by Gabor filter in different orientations I. Introduction In computer vision, the main goal is to develop an automatic/semi-automatic computational technique that takes as input image information and produce as output a particular decision or some form of categorization. The common process entails extracting pertinent features from images, and then feeding these features into a learning algorithm, normally supervised classification. Supervised classification is a machine learning technique that uses labeled training examples to construct a function that can discriminate between discrete classes [1]. Ideally, large quantities of labeled training examples are required in order to create the best classifier model. Some notable applications are such as satellite image classification [2], medical image classification [3], image retrieval [4], [5], and industrial inspection [6]. Texture refers to the visual appearance or actual characteristics of a material’s surface that can be experienced through the sense of touch [1]. Texture can be a useful in computer vision classification tasks since different objects from different classes can exhibit different texture details, allowing strong discrimination capabilities. In this paper, a texture-based approach to classify six Engineering Machined Textures (EMT) into the six classes of Turning, Shaping, Grinding, Horizontal- Milling, Vertical-Milling and Lapping is proposed. A multi-directional Gabor filter is firstly used to extract features from 72-texture images of engineering machined specimens (workpieces). The dimensionality of the resultant feature vector is then reduced using Principal Component Analysis (PCA). Subsequently, this reduced feature set is fed to a supervised Artificial Neural Networks (ANN) for final classification. The main contribution of this work is the use of the Gabor features, which contain texture information from multiple directions. This allows each image to be represented as uniquely as possible, which would further allow accurate discrimination between classes. The rest of this article is organized as follows. Section II provides an overview of related work.