Bengali character recognition using Bidirectional Associative Memories (BAM) neural network M. M. Mahbubul Syeed, Fazlul Hasan Siddiqui, Abu Saleh Abdullah Al-Mamun, Syed Khairuzzaman Tanbeer and M. Abdul Mottalib Department of Computer Science and Information Technology (CIT) Islamic University of Technology, Board Bazar, Gazipur-1704 Emails: rajit_cit@hotmail.com ,fhsani@yahoo.com , mamuncitiut@hotmail.com, tanbeer2000@yahoo.com Abstract: This paper presents the recognition features of Bengali text using BAM (Bidirectional Associative Memories) neural network with a proposal of feature extraction procedure of a Bengali character. To do this, the conventional methods are used for text scanning to segmentation of a text line to a single character. In this paper an efficient procedure is proposed for boundary extraction, scaling of a character and the BAM neural network which increases the performance of character recognition are used. Keyword: Bengali, Character, Neural Network, BAM (Bidirectional Associative Memories), Feature, Scaling, Recognition. 1. INTRODUCTION Bengali character recognition has become an active area of research for last few years with a wide variety of applications. Lots of works already have been done on printed Bengali characters [1] and handwritten Bengali characters [2]. In this paper an approach on recognizing scanned Bengali text is proposed. The Bengali text is scanned first with a scanner and converted into an image format. Then applying several techniques the characters in the text are separated. These separated characters are then applied to a classifier that recognizes the characters as several Bengali characters stored in memory. In this phase BAM (Bidirectional Associative Memories) neural network model is used Neural network is developed to perform some of the activities of human brain. As the recognition of images, voice etc. are performed most efficiently and accurately by human brain, so the artificial neural networks are tried to be developed in computer to perform these recognitions. Like human brain a neural network has a parallel- distributed architecture with a large number of nodes and connections. Each connection points from one node to another and is associated with a weight. A typical model of neural network is shown in figure1. Construction of a neural network involves the following tasks: a) Determination of network properties, as network topology or connectivity, type of connection between the nodes of the network, the order of connection is decided. b) Determination of node properties, as the activation range (discrete [0 & 1] or continuous [0,1]) of a node, type of node activation function (Hard limiting function or Sigmoid function) is determined. c) Determination of system dynamics, as weight initialization scheme, the activation calculation formula, the network learning rule (weight adjustment) is determined. A neural network classifier technique say, BAM is used for recognition. BAM is an associative neural network. An associative neural network is one that retrieves an object or memory based on part of the object itself. The term memory in associative network can be defined as, If a binary n-dimensional vector X is a memory then for each component (neuron) i =1, 2 ,….., n j i Output layer Hidden layer W ji Intput layer Feedforward Connection Recurrent Connection Oj W n W 1 X1 Xn X2 ( ) ∑ = = n i i X i W F j O 1 F= Activation function Figure1: The neural network computational model