IrisIndexNet: Indexing on Iris Databases for Faster Identification

Abstract

Iris is one of the most accurate biometric traits for human authentication due to high reliability, uniqueness, and stability. Identification of a query iris sample requires its comparison with all the templates stored in the database. The process is computationally expensive and becomes inefficient as the number of templates increase in the database. The process can be fastened by reducing the search space through indexing the database. This paper proposes IrisIndexNet, a novel indexing technique that effectively reduces identification search space for the iris database. It design a specialized convolutional neural network architecture and train as a Siamese network to construct a compact feature vectors having low inter-class and high intra-class similarity in the latent space representation for the iris images. Features are subsequently clustered using k-means and agglomerative clustering to generate an index table. During retrieval, the feature of the query image is extracted using the trained network, and it is matched with all the indices of the index table. Feature vectors in neighborhood of most similar index is the candidate set for matching. Obtained candidate set is of fixed size and small as compared to the original database thereby making the identification a constant time operation. Experiments are conducted on widely used publically used iris databases viz. CASIA Interval and CASIA Lamp. The proposed technique achieved a hit rate of 99% at a penetration rate of 2.254% and 0.008% on the respective databases. A speedup of 4 and 27 times is achieved when the CASIA Interval and CASIA Lamp are indexed using the proposed technique as compared to the naive approach for identification.