Combining Subspace Methods and CNN Segmentation for Iris Identification

Abstract

Biometrics provides a wide range of methods for the reliable identification of individuals. Many biometric features are known, but the most reliable among them is the iris texture. It has several advantages, such as uniqueness, durability, stability, collectability and unforgeability. The iris biometric has undergone significant progress in the last few years. Many state-of-the-art methods and approaches are known. This paper presents an iris segmentation and recognition system. The segmentation part is solved by a retrained version of SegNet CNN. It uses the raw image features or Gabor filter responses as input images and applies subspace methods such as PCA and LDA for dimensionality reduction. The final decision in identification is made by a multi-class one-against-one SVM. The performances measured are compared to the CASIA Internal and UPOL databases. The system foreshadows a fusion identification framework applying several types of biometrics.