基于L_(2,1)范式的手部特征融合识别

Hand Feature Fusion Recognition Based on L_(2,1)Paradigm

多模态特征融合是将来自不同模态的特征进行有效整合.传统的典型相关分析的多模态特征融合因没有考虑原子空间中的冗余特征会降低多个特征集的相关性,从而影响融合的效果,因此,提出了基于L_(2,1)范式的耦合特征生物融合识别方法.首先,基于L_(2,1)范式对多个模态的特征集选择耦合特征;其次,基于典型相关分析将多个模态数据集映射到一个公共子空间中,使得投影后的多模态特征集的相关性更佳.从不同的特征空间中选择耦合特征后进行相关性分析,在投影后的子空间中不同模态数据之间具有更好的相似性,从而融合效果更好.基于L_(2,1)范式的耦合特征典型相关生物特征融合的实验结果表明,该方法优于传统的基于典型相关分析多模态特征融合方法.

Multimodal feature fusion is the effective integration of features from different modalities.Traditional multimodal feature fusion based on canonical correlation analysis does not consider that redundant features in atomic space will reduce the correlation of multiple feature sets,thus affecting the fusion effect.Therefore,a coupled feature biometric fusion recognition method based on L_(2,1)paradigm is proposed.Firstly,select coupling features based on L_(2,1)paradigm for feature sets of multiple modes.Secondly,based on canonical correlation analysis,multiple modal data sets are mapped into a common subspace,making the correlation of the projected multimodal feature sets better.This article selects coupling features from different feature spaces and conducts correlation analysis.In the projected subspace,there is better similarity between different modal data,resulting in better fusion results.The experimental results of coupled feature canonical correlation biometric fusion based on L_(2,1)paradigm show that this method is superior to the traditional multimodal feature fusion method based on canonical correlation analysis.