TLDR:

I am disseminating 2 datasets:
Kannada-MNIST dataset: 28X 28 grayscale images: 60k Train | 10k Test
Dig-MNIST: 28X 28 grayscale images: 10240 (1024x10) {See pic below}

The Kannada-MNIST dataset is meant to be a drop-in replacement for the MNIST dataset 🙏 , albeit for the numeral symbols in the Kannada language .
Also, I am disseminating an additional dataset of 10k handwritten digits in the same language (predominantly by the non-native users of the language) called Dig-MNIST that can be used as an additional test set.

GitHub 👉: 
Kaggle 👉: 
ArXiv 👉 : 

Prabhu, Vinay Uday. “Kannada-MNIST: A new handwritten digits dataset for the Kannada language.” arXiv preprint arXiv:1908.01242 (2019)..

@article{prabhu2019kannada,

title={Kannada-MNIST: A new handwritten digits dataset for the Kannada language},
  author={Prabhu, Vinay Uday},
  journal={arXiv preprint arXiv:1908.01242},
  year={2019}
}

Introduction:

Fig: A vehicle license plate with Kannada numeral glyphs

This figure below captures the MNIST-ized renderings of the variations of the glyphs across the following modern fonts: Kedage, Malige-i, Malige-n, Malige-b, Kedage-n, Malige-t, Kedage-t, Kedage-i, Lohit-Kannada, Sampige and Hubballi-Regular.

Dataset curation

Kannada-MNIST:

Volunteers helping curate the Kannada-MNIST dataset

Dig-MNIST:

Comparisons with MNIST:

Some classification benchmarking:

I used a standard MNIST-CNN architecture to get some basic accuracy benchmarks (See fig below)

The CNN architecture used for the benchmarks

(a) Train on Kannada-MNIST train and test on Kannada-MNIST test

(b) Train on Kannada-MNIST train and test on Dig-MNIST

Open challenges to the machine learning community

1. To characterize the nature of catastrophic forgetting when a CNN pre-trained on MNIST is retrained with Kannada-MNIST. This is particularly interesting given the observation that the typographical glyphs for 3 and 7 in Kannada-MNIST hold uncanny resemblance with the glyph for 2 in MNIST.
2. Get a model trained on purely synthetic data generated using the fonts (as in [1]) and augmenting to achieve high accuracy of the Kannada-MNIST and Dig-MNIST datasets.
3. Replicate the procedure described in the paper across different languages/scripts, especially the Indic scripts.
4. With regards to the dig-MNIST dataset, we saw that some of the volunteers had transgressed the borders of the grid and hence some of the images either have only a partial slice of the glyph/stroke or have an appearance where it can be argued that they could potentially belong to either of two different classes. With regards to these images, it would be worthwhile to see if we can design a classifier that will allocate proportionate softmax masses to the candidate classes.
5. The main reason behind us sharing the raw scan images was to foster research into auto-segmentation algorithms that will parse the individual digit images from the grid, which might in turn lead to higher quality of images in the upgraded versions of the dataset.
6. Achieve MNIST-level accuracy by training on the Kannada-MNIST dataset and testing on the Dig-MNIST dataset without resorting to image pre-processing.

[1] Prabhu, Vinay Uday, Sanghyun Han, Dian Ang Yap, Mihail Douhaniaris, Preethi Seshadri, and John Whaley. “Fonts-2-Handwriting: A Seed-Augment-Train framework for universal digit classification.” arXiv preprint arXiv:1905.08633 (2019). [  ]