| Abstrakt: |
Recent literatures highlight the potential of high dimensional features in designing representations in object recognition. Binarized Statistical Image Features, dubbed as BSIF, is a high dimensional representation and advocates its capability in texture classification and face recognition. Borrowing the idea of BSIF, we utilize ICA filters learnt from a set of natural training images to construct a new unsupervised learning technique, namely Histogram of Independent Component Pattern, coined as HICP. HICP further consolidates ICA response invariance through binarizing the filter response into a binary map and block-wise histogramming the outputs. The binarization and block-wise histogramming allow nonlinear operation in HICP which further boost the discriminating capability. Besides, a signed square root normalization operation on HICP features suppresses those numerically dominating entries trigged by zero padding in the block-wise histogramming process, particularly at the cell boundaries padded with zero. We evaluate the recognition performance of HICP on face recognition under several scenarios, comprising of facial expression variation, illumination variation, time span and facial makeup effect. The empirical results demonstrate that the proposed HICP is able to achieve on par with or even better recognition performance than the other existing state of the art techniques. [ABSTRACT FROM AUTHOR] |
