脸耳多模态稀疏表示融合识别方法比较

将稀疏表示应用于脸耳多模态身份辨识，比较和分析采用不同融合方法的多模态稀疏表示识别算法的准确性和鲁棒性，为多模态稀疏表示融合识别算法设计提供理论和方法指导。结合多模态融合层次理论与稀疏表示分类的技术特点，提出3种多模态稀疏表示识别方法:直接特征融合法、间接特征融合法和匹配层融合法。从多模态融合角度看，3种方法的不同在于融合层次或融合策略不同；从稀疏表示角度看，它们的主要区别在于稀疏表示时脸和耳特征耦合的程度不同。在3个多模态数据库上的实验结果表明:所提3种方法在识别准确率和鲁棒性上远优于采用NN、NFL和SVM等分类器的融合识别方法；当脸耳图像中噪声不显著时, 3种方法性能相当，当噪声严重时，匹配层融合识别方法优于特征层融合方法。

Comparison of Face and Ear Based on Multimodal Biometric Identification with Sparse Representation

This paper proposes to employ sparse representation (SR) in multimodal biometric identification of face and ear, and focuses on performance comparison among the presented approaches with different fusion schemes seeking to find guideline for designing mulitimodal biometric recognition systems with sparse representation. In this paper, three multimodal methods are introduced based on the hierarchical multimodal fusion theory and SR's operating mechanism. These methods are MSRCef (multimodal SRC with explicit feature fusion), MSRCif (multimodal SRC with implicit feature fusion), and MSRCs (multimodal SRC at score level). From the viewpoint of multimodal fusion, they adopt different fusion strategies, on the other hand, their major difference lies on the constraint imposed on the sparse representation of face and ear features. Experimental results on three multimodal databases demonstrate that all the three proposed multimodal approaches perform significantly better than those using NN, NFL, SVM, etc. Besides, the proposed multimodal methods are generally comparable, however the method with score level fusion scheme is obviously superior to the others with feature level fusion when the face and/or ear images confront heavy corruption.