A New Approach of Sign Language Recognition System ... - IEEE Xplore

4 downloads 11033 Views 1MB Size Report
1 st International Conference on Electrical & Electronic Engineering (ICEEE) ... simple, low cost Bangia sign language translation (BSLT) system that can ...
1st International Conference on Electrical & Electronic Engineering (ICEEE) 04-06 November 2015, RUET, Rajshahi, Bangladesh

A New Approach of Sign Language Recognition System for Bilingual Users S. M. Kamrul Hasan \ Mohiuddin Ahmadl,2 IDepartment of Electrical and Electronic Engineering, 2Department of Biomedical Engineering Khulna University of Engineering & Technology (KUET) Khulna-9203, Bangladesh Email: [email protected]@gmail.com

Abstract- Sign language is becoming increasingly popular day by day to make a bridge between the hearing impaired and normal people. It is a very challenging task in respect of developing country like Bangladesh where around 2.4 million people use Bangia sign language. In this respect we propose a simple, low cost Bangia sign language translation (BSLT) system that can translate sign into BangIa text. We report on the development of universal interpreter software (UIS) that can be used by both the American and the Bangladeshi users. For this, an efficient method is proposed for skin detection & feature extraction. Our system can recognize 16 Bengali words & 11 Bengali numbers. We train our system using a database of (27xl0x20) images, i.e. 10 persons containing 20 images per sign & for testing we use another 2700 (27xlOxlO) images. The system

To overcome the impediments & to increase accuracy, we propose a simple, low cost BSLT system using a new approach that combines both Principal Component Analysis (PCA) & Linear Discriminant Analysis (LDA) for feature extraction & also dimension reduction. It overcomes the limitations of PCA. The most attractive feature of our system is to use the same database for both the Bangladeshi & the American users. The system can successfully recognize 16 daily necessary words & 10 BangIa numbers (0-':>0). The outline of this paper is as follows: From Section II-IV proposed BSLT system has been depicted. Section V focuses on results & conclusion is drawn in section VI.

results in about 96.463% accuracy as compared to K-Nearest

II.

Neighbor algorithm.

PROPOSED BANGLA SIGN LANGUAGE TRANSLATION

(BSLT) SYSTEM Keywords- Bangia sign language translation (BSLT) system; eigen vectors; fisher vectors; Support Vector Machine; universal interpreter software (VIS).

I.

INTRODUCTION

With the advancement of technology, life has become easier and comfortable. But, to avail this technology is a big issue for developing country like Bangladesh. Here, there is a huge communication gap between the hearing impaired & the normal people. So, sign language can be a subsidiary medium which is a visual form of language, including hand gesture or the facial expression. Bengali Sign Language (BdSL) is different from other countries. So, not a more researchers attempt at BdSL recognition system. In general, most attempts either relied on data gloves or most used single algorithm to recognize sign language. S. Begum & M. Hasanuzzaman use Principal Component Analysis (PCA) for both preprocessing & recognition [1]. Due to absence of improved classification algorithm they could not get much accuracy. B. C. Karmokar, K. M. K. Alam & M. K. Siddiquee recognize BdSL using Neural Network Ensemble. But they didn't use any recognized algorithm for preprocessing even they didn't clarify the preprocessing method [2]. Zhang & Qiao [3] apply both linear and nonlinear SVM training models for recognition purpose. But during training the SVM, the whole images are used as training data that requires an expensive calculation. Therefore, the authors take a decision to carry out feature extraction in the further research.

978-1-4673-7819-2/15/$31.00©2015 IEEE

Our proposed system needs no wearable gloves and so it is user friendly. We divide our work into five stages: Image Acquisition, Preprocessing, Feature Extraction, Sign Classification & Translation in BangIa Text using a GUI (Graphical User Interface). The input of BSLT system is the hand sign images for testing, recorded after a pause of 5 seconds using a webcam. To train this system we create a dataset of 27 different Bangia word signs each of 50x50 pixels from 10 different. All the training signs are the benchmark dataset belonging to "Deaf cultures: Bangladesh" section.

Fig. 1.

Testing dataset for BangIa sign words(left to right & top to bottom)

�,�, m ,mPIT, 'm'fr,

�,�,�,� ,

Suggest Documents