Abstract:
Creating automatically handshape classification inventory is a time-consuming process,
in view of the fact that handshape datasets have to be carefully classified by
linguists. Thus, only some of the popular languages have such handshape automatic
classification inventory.
This project aims to create a strong algorithm to classify a large unlabeled dataset
of sign language handshapes. Previous works in image classification show significant
results of more than 80% accuracy, but there are no relevant sources that echoed the
same results in the classification of large weakly-labeled handshapes dataset using
sem-supervised learning. As it was mentioned previously, the dataset is one of the
main problems in that theme. In this work, we have a large set of unlabeled samples
and about 45 classes of labeled image samples. Therefore, the selected approach
should work well even when labeled data are not abundant. It is planned to test
the semi-supervised learning approaches that take advantage of the small but labeled
set such as Noisy-Student Training and it s expected to outperform results of the
supervised Deep Hand model on the same dataset.