Which Dataset is this Iris Image From?

摘要

The performance of a biometric recognition algorithm is often evaluated by testing it on standard datasets. This process, known as technology evaluation, is necessary to compare the matching performance of different algorithms. In the case of iris recognition, datasets such as ICE, MBGC, CASIA, NICE, WVU, UBIRIS, etc. have been used for this purpose. However, iris images in each of these datasets are impacted by the methodology used to collect them. Factors such as external lighting, sensor characteristics, acquisition protocol, subject composition, data collection environment, nuances of the collection process, etc. are dataset-specific and they leave a digital 'imprint' on the associated data. Therefore, iris images in different datasets may exhibit different intricate characteristics that can potentially impact the performance assessment process. In this work, we conduct an experiment to determine if such dataset-specific attributes are significant enough to be detected in the collected images. To this end we formulate a classification problem where the goal is to determine the dataset to which a given input iris image belongs to. By extracting a set of statistical and Gabor-based features from an iris image, we use a learning-based scheme to associate the input iris image with a specific database. A 83% accuracy is obtained on a set of 1536 images from 8 different datasets collected using 6 different sensors.