--

This project was a challenge to myself to write a functioning OCR program using contemporary technology. It is a testament to the programmer productivity afforded by the python language and available libraries and documentation that it took only eight person-days of concerted effort to achieve this. Decades ago, the language of choice might have been C. There was no stack exchange or even google, let alone the world-wide web, just photocopy machines, microfiche readers, and card-catalogs for research, magnetic tapes or dial-up modems for file transfer, and most computing chassis weighed as much or more than an adult human. I might estimate that duplicating the functionality of that 8 day effort a few decades ago would have taken the order of a year, even if all of the algorithms were known.

It was also a beginner project in aspects of machine learning and various python libraries.

A cellphone camera is used to capture images of printed data. The original document was a numerical table printed in a journal. The set of characters of interest is small (i.e. the digits 0-9 and numerical signs and punctuation). The image is warped due to page curvature and suffers from moderate blurring.

It is strongly desirable that some indication is given to the user of the certainty of classification of each character glyph, so that the error rate of the entire process (imaging, training the classifier, and post-editing the output) can be made arbitrarily small, given an image that is not too blurry or warped. The amount of required post-editing should be minimized.

The possible categories to be predicted (classes) can be arbitrary strings if it desirable to recognize characters in different fonts.

• feature_sel.py extracts features (actually instances - character glyphs) from an image.
• cluster.py performs hierarchical clustering of the character glyphs and interactively prompts the user for training data for the specific fonts inherent in the image.
• classify.py classifies the remaining character glyphs, estimates character spacing and layout, performs left-to-right text flowing and output. (actually the classifier currently just uses the results of the earlier clustering phase)

• Satisfactory deblurring the image is desirable. Some form of Bayesian analysis could be used to estimate a usable Point Spread Function for use as a deconvolution kernel. Some attempt has been made using the pymc probabilistic programming suite. The model specification needs help. STAN might be preferable to pymc.
• Review use of terminology throughout the project.
• Provide GUI for entering training data.
• Implement some sort of neural network with saved training set for common generic feature sets as an alternative to manual training for specific fonts
• Explore what, if anything, this project might provide to other projects such as ocropus/kraken.
• Process a montage of images of the same source page, handling overlaps, different warps, skews and blurrings. Maybe Bayesian modelling of it all.