OCR Development- Bossonova Project

The most common ways for the OCR baseline was the Convolutional Recurrent Neural Networks. Even though, it was very effective for letters, it did not work well for the price tags.

As we can see the size of the numerical characters are bigger than the normal text, hence it was difficult to detect it using tesseract detector as well.

The method we came up with is a very small model with very few parameters:

Using the above pipeline, we were able toachieve test accuracy of 100% with test recall of 95.65%