Handwritten Sylheti Nagri Letters: A 32-Class Image Dataset for Letter Recognition

A total of 12,800 labelled images are present with each class having 400 images. The hand-written samples were obtained with paper sheets from over 400 participating university student volunteers, which allowed for variation in writing style, letter shape, stroke thickness, letter size, and letter orientation.

Once the data was collected, the qualified sheets were chosen for clarity and usability of the hand written samples. A custom semi-automated image-separation tool was then used to extract individual letter images from the sheets and organize them according to their respective classes. Finally, the entire set of data was checked manually to ensure that the images were properly separated, labelled and classified in respective class folders.