Bangla Sign Digits Recognition Using HOG Feature Based Multi-Class Support Vector Machine

摘要

Sign language or gesture language is used by the people who are unable to hear for communicating among themselves as well as with other persons. It is not as simple as the natural speaking language. It is a fully developed natural language having its grammar and lexicon. Gesture Languages are represented via the manual sign-stream or alphabet-stream together with non-manual components. As the spoken languages have various types of sounds, letters, digits, etc. Similarly, sign languages have their grammar, letter, digits, etc. In this paper, the main focus is sign digits as the sign numerical digits are a major part of a sign or gesture language which are not familiar to general people. So the goal is to make it understandable to the general people. For this purpose, an easily understandable model has been constructed with computer vision features and machine learning methods to recognize Bangla finger numerical digits. The histogram of oriented gradient features of images has been applied to train the classifier, here a multiple-class support vector machine has been employed to classify the images. In this paper, the proposed multiple-class Support Vector Machine is trained with the HOG features of the images with the training images dataset to achieve the goal. The classifier is trained and tested for 900 training pictures and 100 test pictures of ten-digit classes respectively. After training and tests, the proposed classifier model has gained about 95 percent accuracy on the perception of Bangla sign numerical digits.