Skip to content
Valerii Zuev edited this page Feb 12, 2021 · 11 revisions

Braille OCR

Objective

Angelina Dataset (see "links" section) consists of 240 labeled images with Braille text (colored, 96 dpi, usually 1024x1344 or 1024x1376 pixels), divided into a train set (212 pictures) and a validation set (28 pictures). Braille text is written in a 6-dot code, each letter is formed by 1 to 6 raized dots located in a 2*3 cell. A label corresponds to a single character; the character location is defined by two points (its upper left and lower right corners). An image may contain recto as well as verso dots; only recto dots should be detected.

Double-sided Braille book Single-sided Braille writing
Double-sided Braille book Single-sided Braille writing

The primary goal is to build a detector that recognizes Braille characters on validation images with high accuracy (in accordance with the IoU loss), presumably using the training dataset.

The next goal, which I will aim only if I succeed in the first one, is to add more images to the Angelina Dataset which contain Braille characters not only from books and Braille paper sheets but also those from Braille plates and other materials such as Braille Contraction cards used by students; then build a new detection system that will perform well on the new data, too.

In addition, I plan to combine the dataset with unlabeled images that I will shoot at home, in the Library for the Blind, Rehabilitation Center For The Blind, etc., in order to improve the detection performance using semi-supervised learning (e.g. semi-supervised data mining).

Links

Braille OCR

Semi-Supervised Learning

Papers

Braille OCR

Semi-Supervised Learning

Miscellaneous

Clone this wiki locally