Siamese-CNN for comparing prehistoric and early historic coins

This is a project which was part of my University of Surrey Comp Sci degree. I applied a siamese CNN model to compare extracted image features of some 7000 images scraped from the British Museum Catalogue.

Preparing the Data

• To begin we need to scrape our image data by running BMScraper.py and then imageDownloader.py.
• Some images might have a black background, there's enough data to justify removing them but some may be considered valuable, if so run invert.py on the coin group.
• We apply histograms to the image data by running imageHist.py, you may choose to ignore this step or modify the histogram being used.
• Finally the data needs to be converted to a MNIST CSV format, run mnistMaker.py.

Training the Model

• We train two models, feature and simmillarity. Feature compares groups and Simmilarity compares two images, run siamese.py.
• To display the results run loadResults.py

Authors

• Charlie Tizzard - Initial work

Acknowledgments

• Colin O'Keefe for the Web Scraper