Published: 16 May 2023| Version 1 | DOI: 10.17632/6f2wm5p3vf.1
Nihal Md Ragib Amin,


the existing datasets offer very limited variety in terms of background, light contrast, skin tone, capture angle, number of subjects, and image scaling. As the development of an efficient learning model to deal with real-world scenarios requires a large variety of data, we endeavored to create a large dataset of one-handed BdSL alphabet. We gathered 35,149 images of 37 one-handed BdSL signs combining different previously introduced datasets and labeled the images into 37 classes where each class possesses 950–1000 images. The images have over 150 types of background and a broad range of light contrast, hand size, image scale, and skin tone of hand. The images are captured from more than 350 subjects and various angles. BdSL-MNIST is a newly developed dataset consisting of 64x64 pixel images. BdSL-MNIST provides a valuable resource for researchers and developers working on sign language recognition systems for Bengali-speaking individuals.



University of Dhaka Faculty of Engineering and Technology, Tokyo Kogyo Daigaku, Meiji Daigaku


Image Classification, Sign Language, Bengali Language, Recognition