Senior Honors Thesis: "A Comparison of Iris Scanning Algorithms: Classical Vs. Machine Learning."

Abstract: Humans have long used biometric identification methods to identify others. One recent method is scanning the iris, which is known to have unique patterns similar to the fingerprint. Unlike the fingerprint, irises can be examined at a distance without needing to contact a surface. Coinciding with the growing popularity of iris scanning is the growing popularity and efficacy of machine learning. As a result, many researchers have used machine learning models to determine whether two iris photographs are from the same individual or not. However, it is not well known whether these novel machine learning algorithms outperform their classical counterparts in speed, accuracy, or performance given low-resolution data.

To compare the two approaches, a classical algorithm and machine learning algorithm for iris scan matching are implemented in Matlab and Python. Both algorithms use the original CASIA Iris V3 dataset and a 5x blurred version of it. A confusion matrix is produced for each algorithm and each data set. Each algorithm is then given one half of the CASIA Iris V3 dataset as a “population” and a new iris scan. The algorithms are asked to determine who, if anyone, in the population matches the new scan. The classical algorithm outperforms the machine learning algorithm in performance in both data sets and both tests but has a longer runtime. Finally, the CASIA dataset is further blurred by factors beyond 5x, and a factor of 17.5x is identified as the machine learning-superior blur factor. The machine learning algorithm outperforms the classical algorithm at blur factors greater than 17.5x and underperforms otherwise. Despite the machine learning algorithm being generally faster than the classical version, I conclude that the classical implementation’s accuracy is so far superior that it is the obvious choice for any real-world implementation.

Host: Victor Wickerhauser