Signal-Level Information Fusion for Less Constrained Iris Recognition Using Sparse-Error Low Rank Matrix Factorization

Iris recognition systems working in less constrained environments with the subject at-a-distance and on-the-move suffer from the noise and degradations in the iris captures. These noise and degradations significantly deteriorate iris recognition performance. In this paper, we propose a novel signal-level information fusion method to mitigate the influence of noise and degradations for less constrained iris recognition systems. The proposed method is based on low rank approximation (LRA). Given multiple noisy captures of the same eye, we assume that: 1) the potential noiseless images lie in a low rank subspace and 2) the noise is spatially sparse. Based on these assumptions, we seek an LRA of noisy captures to separate the noiseless images and noise for information fusion. Specifically, we propose a sparse-error low rank matrix factorization model to perform LRA, decomposing the noisy captures into a low rank component and a sparse error component. The low rank component estimates the potential noiseless images, while the error component models the noise. Then, the low rank and error components are utilized to perform signal-level fusion separately, producing two individually fused images. Finally, we combine the two fused images at the code level to produce one iris code as the final fusion result. Experiments on benchmark data sets demonstrate that the proposed signal-level fusion method is able to achieve a generally improved iris recognition performance in less constrained environment, in comparison with the existing iris recognition algorithms, especially for the iris captures with heavy noise and low quality.