Multimodal feature fusion is the effective integration of features from different modalities.Tradi-tional 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 L2,1 paradigm is proposed.Firstly,select coupling features based on L2,1 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 L2,1 paradigm show that this method is superior to the traditional multimodal feature fusion method based on canonical correlation analysis.
万方数据