This is team NLP's final project for CPE 428. For this project we did handwritten OCR using a convolutional neural network trained on a modified version of the Chars74k dataset.
The neural network implementation was done using Tensorflow and was based off of this Tensorflow wrapper.
We made major additions to the Tensorflow wrapper for the purposes of this project. This includes adding the ability to randomly expand the dataset between each training iteration.
To save the neural network to a file for later use we have to use the improved pickling library dill to do so. Standard python pickling is unable to pickle certain tensorflow variables while dill can do so.
For word recognition we made extensive use of numpy and scikit-image to segment the letters out of the given image.
The source code for word recognition can be found in "ocr_util.py" which is somewhat of a misnomer. It is not so much utilities, it quickly evolved into our word recognition implementation.