Purchase this article with an account.
Sandro De Zanet, Michael Rueegsegger, Tobias Rudolph, Sebastian Wolf, Jens Kowal; Evaluation of Feature Detectors and Descriptors on the Iris. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5412.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Characterizing and matching human iris patterns is essential for video tracking in ophthalmic surgery as well as in iris recognition. This study presents an evaluation of feature detection and description algorithms. The evaluation is based on a database of 19 eyes captured in the infrared spectrum.
Repeatability and precision/recall were measured for combinations of state-of-the-art feature detectors and descriptors: SURF, SIFT, FAST, STAR and BRISK as detectors and SURF, SIFT, FREAK, BRISK and BRIEF as descriptors. The algorithms were tested with images of the iris in different geometric and photometric transformations. A ground truth was provided by using manually segmented and non-rigidly registered irides and masking out reflections. Images were acquired from 40 human irides. The eyes were illuminated by infrared LEDs and captured by a camera via a beam splitter on a slit lamp microscope. To fixate the gaze of the patients yellow LEDs were used to indicate different gaze angles. Images where a good ground truth could not be found were discarded which results in 19 testable irides.
The most accurate detector/descriptor combination has been found to be STAR/SU-BRISK which features fast detection in combination with fast matching through the Hamming distance. Especially the SU-BRISK descriptor has proven to be the most distinctive independent from the used detector. Generally orientation- and scale-variant descriptors performed far better than their invariant counterparts. Additionally binary descriptors based on intensity comparisons like BRIEF, BRISK or FREAK yield a more discriminatory descriptor than real value based vectors. Through the evaluation of detected point distributions it is evident that the region from the collarette to the pupil border is the are with the most features. Gaze change of 15° proved to decrease the matching accuracy. However, decreasing the contrast with less intensive light impaired accuracy more than gaze change. The area under the precision-vs-recall-curve shows a range of 0.01 to 0.71 which indicates the importance of isolating the best configuration for iris images.
In this study a good candidate for real-time feature-matching has been found which can be used in torsional eye tracking used in beam therapy, namely the STAR detector with the scale- and rotation-variant BRISK descriptor.
This PDF is available to Subscribers Only