A novel facial feature extraction method based on Empirical Mode Decomposition

Abstract

Empirical Mode Decomposition(EMD) is a signal decomposition technique for adaptive representation of signals, as the sum of a set of Intrinsic Mode Functions(IMFs). It captures signal information that contains local trends by measuring signal oscillations, which can be quantized by some local high frequency components or local low frequency components, i.e., IMFs. Orthogonality of the IMFs is an important index to measure the performance of the EMD method. However, instead of elaborating theoretically, most literatures only check the orthogonality in practical sense. In other words, the direct IMFs are not exactly orthogonal. In this paper, in order to get orthogonal IMFs, we orthogonalize the IMFs. According to the sequence, two ways of orthogonalization can be developed: one is from the top IMF to the last, and the other is in the contrary sequence. The orthogonal IMFs can express the original signal more accurately. For the first orthogonalization strategy in face recognition, namely from the top IMF to the last, they are sorted from the highest frequency to the lowest frequency, which correspond to the detail of coarse facial information (facial features). In our experiments, several strategies are designed to find the most efficient orthogonal IMF. Excellent performances are obtained on the eMU PIE database.