Variable Length Digit Recognition for Gujarati Language

In this paper, we describe a method to perform handwritten digit recognition for Gujarati - a regional Indian language. Our method can handle variable-length inputs, meaning that there are no limitations around the digit length for the input image. To our knowledge, this is the first attempt to do variable length digit classification for the Gujarati language numerals. We outline a two-step method to classify handwritten Gujarati numerals. The first step identifies connected components of the input image and predicts the numeric length of each connected component. The second step predicts the actual number that is contained within each connected component. The final result is a concatenation of individual predictions. Our Convolutional Neural Networks (CNN) architecture for this task has a low number of output classes (e.g. 30 classes for 3 digit classifier). Our method achieves 83.8% test set accuracy for 1 to 4 digit Gujarati numerals. On the NIST19 dataset, our method achieves 96.1% test set accuracy for 2 to 6 digit English numerals.