This repository is used to convert images of ascii characters to text using tensorflow.
Characters can be one of 94 classes - [0-9][A-Z][a-z] and symbols such as (+-./| etc.
Images must be a size of 64x64 px - Images of different sizes should be resized to 64x64, not padded. Padding does not work so well if the models are trainined using the randomly generated dataset. Images should have exactly one character in them.
This OCR is implemented using Tensorflow 2.x.
There are a number of pre-made models which can easily be trained on a generated dataset.
The file utils/img_to_dataset.py can be used to convert the chars74k data (Download link) to a Tensorflow dataset able to be used with the project.
Images can also be manually synthesized with data_generation/create_images.py. This will save the data to the data directory by default. You can specify whether you want to save the generated images as raw .png files, or as Tensorflow datasets which are automatically saved and can be loaded to train the models with.
Here are some examples of the randomly generated images:
The file utils/manually_classify.py is provided for you to have a go at manually classifying these images to get an idea of human level performance.