Alphanumeric Character Recognition on Tiny Dataset

Abstract

Alphanumeric character recognition is the ability of a computer to understand digits and alphabets image as an input. This topic has struggled to reach a wide variety of people since we need quite a lot of labeled training data. A small training sample achieves sub-standard accuracy in training data that is model is not able to learn accurately. Because of this handwritten (alphanumeric character), recognition is not widely been used. We resolve this issue by creating new labeled training data. Generated data is generated from existing samples, with a realistic data expansion method that is an actual variant of human handwriting. This is done by changing or adding more value to the parameters. Our model had just 10–11% training samples per character compared to other character-recognition methods. We achieved close to existing state-of-the-art character recognition results of all four datasets. Images reconstruction strategy was also used. Our system can be used when training data is less, and the user wants to achieve reasonable accuracy.