Robust multimodal recognition via multitask multivariate low-rank representations

We propose multi-task, multivariate low-rank representation-based methods for multimodal biometrics recognition. Our methods can be viewed as a generalized version of multivariate low-rank regression, where low-rank representation across all the modalities is imposed. One of our methods takes into account coupling information among different biometric modalities simultaneously by enforcing the common low-rank representation within each biometric's observations. We further modify our methods by including a background occlusion term that is assumed to be sparse. Alternating direction method of multipliers is proposed to solve the proposed optimization problems. Extensive experiments using face and touch gesture dataset show that our method compares favorably with other feature level fusion-based methods.