SIFT based approach on Bangla sign language recognition

This paper presents a SIFT-based geometrically computational approach to vigorously recognize Bangla sign language (BdSL). Gaussian distribution and grayscaling techniques are applied for image processing and normalizing the sign image. After this pre-processing, features are extracted from the sign image by implementing scale invariant feature transform. Acquiring all descriptors from the sign image, k-means clustering is executed on all the descriptors which are previously computed by SIFT. Based on the sample training set, each of the cluster denotes as a visual word. Considering the histograms of the clustering descriptors, Bag of words model is introduced on this hybrid approach which develops a set of visual vocabulary. Finally for each of sign word, a binary linear support vector machine (SVM) classifier is trained with a respective training data set. Considering these binary classifiers, we obtained a respective recognition rate on both Bangla signs of expressions and alphabets.