This thesis evaluates the impact of eye diseases on biometric verification, considering the drop in accuracy of recognition they can cause and the concerns about the privacy of the individuals. Therefore the database of the Warsaw University of Technology, which contains images of sick eyes, and the Iris Toolkit of the University of Salzburg are used. Both are freely available which makes the work completely reproducible. We too want to know if sick eyes affect the template protection schemes. To that end, we use a Bloom filter based algorithm. As a second test the database was used to classify healthy and sick eyes with Support Vector Machines. The python scikit implementation of Support Vector Machines was used. This is also freely available so the whole thesis is reproducible.
Eye diseases have a large impact on iris recognition. The experimental results have shown that even clear, sick eyes have an accuracy drop of ten percent compared to that of healthy eyes. Compared to general healthy results that would mean a accuracy drop on hundred percent. For visibly changed eyes the drop is even higher. Bloom filter based template protection can be applied without a noticeable change in recognition accuracy with respect to the unprotected systems.
In a second set of experiments, the feasibility of distinguishing sick from healthy eyes was tested. With a Support Vector Machine we used the Local Binary Pattern features from the database images to classify sick and healthy eyes. Here we used the visibly sick eyes as training data, because Local Binary Pattern are a visual descriptor.
The results show, sick eyes have similar colour gradient the Support Vector Machine either has a really high false match rate for healthy eyes and recognizes disease that have little visible change or has a really high false match rate for sick eyes, but recognizes healthy eyes right.