NSU Research Contributions
Title : Bangla Sign Language Alphabet Recognition Using Transfer Learning Based Convolutional Neural Network
Authors : Kanchon Kanti Podder, Muhammad E. H. Chowdhury, Zaid Mahbub, Muhammad Abdul Kadir
Abstract : Sign language is the preferred medium of communication for the hearing and speech impaired people. Bangla sign language (BdSL) recognition systems and their practical implementations are still at their infancy. This paper describes a convolutional neural network (CNN) based deep transfer learning approach using ResNet18 for the interpretation of all the 37 static signs of the one-handed Bangla sign alphabet. A color-coded fingertip pattern was implemented to minimize inter-class similarity of sign gestures. A large dataset, consisting of 45,958 images of all the sign letters (ranging from 821 to 1999 images for each sign) collected from 5 different subjects, has been created as part of this work and made publicly available. The transfer learning model was trained and validated using five-fold cross validation for 37-class classification and then the performance in recognizing BdSL signs were evaluated. Experimental results showed that the method can recognize Bangla signs with an overall accuracy, sensitivity and specificity of 99.97, 99.49 and 99.99 %, respectively. The proposed method of automatic BdSL recognition with very high accuracy can aid the hearing and speech impaired children in learning sign language as well as help millions of sign language users in their day-today communication.
| Journal : Bangladesh journal of scientific research | Volume : 6 | Year : 2020 | Issue : 1 |
| Pages : 20-26 | City : | Edition : | Editors : |
| Publisher : | ISBN : | Book : | Chapter : |
| Proceeding Title : | Institution : | Issuer : | Number : |