Automatic recognition of Turkish fingerspelling

Fingerspelling is a manual representation of alphabet letters and used in sign language to spell out words, especially private names. In Turkish Sign Language it is a challenging task due to the ambiguity of the fingerspelling representations of the letters with diacritic marks and complex hand configurations. In this paper we propose a Turkish fingerspelling recognition system based on extraction of the most effective features that help disambiguation of the signs. Our approach is fundamentally based on feature extraction by Histograms of Oriented Gradients (HOG) and dimension reduction by Principal Component Analysis (PCA). Use of internal features instead of outer projection of the image enhances the performance on disambiguation. The test dataset consists of 493 fingerspelling images created by 4 different signers and the test results indicate an average success rate of classification of 99.39%.