Tensor decomposition of SIFT descriptors for person identification

This paper studies person identification using human iris based on tensor decomposition of SIFT features. First, we divide a normalized iris image into small non-overlapping image patches, each of which is represented by a SIFT descriptor. In this way, the iris image is naturally represented by a fourth-order tensor. We use tensor decomposition to obtain features of reduced dimensionality by the alternating least square algorithm. The low-dimensional features are encoded to binary codes by comparing with the mean value of every dimension. We perform iris matching by counting the average number of two binary codes in agreement. In the iris matching process, the occlusion or noisy factors are also considered. The proposed method is validated with the UBIRIS.v2 and CASIA-IrisV4 datasets.