The IIIT-Synthetic-M-Hindi dataset consists of synthetically created 2,78,248 word images along with their corresponding annotations. To create synthetic images, freely available Unicode fonts are used to render synthetic word images. The number of unique fonts used for Hindi is 97. We use ImageMagick, Pango, and Cairo tools to render text onto images. To mimic the typical document images, we generate images whose background is always lighter (higher intensity) than the foreground. Each word is rendered as an image using a random font. Font size, font styling such as bold and italic, foreground and background intensities, kerning, and skew are varied for each image to generate a diverse set of samples. A random one-fourth of the images are smoothed using a Gaussian filter with a standard deviation (𝜎) of 0.5. Finally, all the images are resized to a height of 32 while keeping the original aspect ratio. This dataset can be used only for pre-training purposes. There are 24,604 unique Hindi words in the training set.

Training Set:

train.zip contains folder named “images” with 2,78,248 word level images, “train_gt.txt” containing image name and ground truth text separated by “Tab space” and and “list_of_unique_words.txt” contains list of 24,604 unique words in the Training set.