IRIS RECOGNITION ON HADOOP

Abstract

Iris Recognition is a type of pattern recognition which recognizes a user by determining the physical structure of an individual's Iris. A unique Iris pattern is extracted from a digitized image of the eye, and encoded into an iris template. Iris template contains unique information of an individual and is stored in a database. To identify an individual by iris recognition system, an individual’s eye image is captured using video camera and converted into iris template. These templates are compared with stored iris templates in database. If templates are matches then user is said to be genuine, otherwise imposter. Iris Recognition offers advantages over traditional recognition methods (ID cards, PIN numbers) because the person to be identified has no need to remember or carry any information. Iris pattern remains stable throughout life of a person. This characteristic makes it very attractive for use as a biometric for identifying individual. Iris recognition is deployed for verification and/or identification in applications such as access control, border management, and Identification systems. With increasing security concerns, biometric database size is growing very fast and technologies like iris recognition has a very huge database for comparison. Iris recognition algorithms are implemented on general purpose sequential processing systems, and also existing relational database systems are not enough to handle this huge size of data in some reasonable time. In this thesis, a parallel processing alternative using cloud computing, offering an opportunity to increase speed and use it on huge database is proposed. An open source Hadoop framework for cloud computing is used to implement the proposed system. Hadoop Distributed File System (HDFS) is used to handle large data sets, by breaking it into blocks and replicating blocks on various machines in cloud. Template comparison is done independently on different blocks of data by various machines in parallel. Map/Reduce programming model is used for processing large data sets. Map/Reduce process the data in <key, value> format. Iris database is stored in a <key, value> text format. Mappers process the input and produce an intermediate output. Reducer takes intermediate output and produces final result. This research work shows how, the most time-consuming operations (matching process) of a modern iris recognition algorithm are parallelized. In particular, template matching is parallelized on a cloud based system with a demonstrated speedup gain.