Description

ABSTRACT:

          we propose an approach named the local energy pattern (LEP), which can be considered as an extension of LBP. The main contribution is two-fold, corresponding to the two fundamental problems of statistical texture representation. First, is the generation of the local feature vector. The normalized local oriented energies which are obtained by rectifying the responses of the 2D Gaussian-like second derivative filters are used to generate the local feature vectors. These local feature vectors are invariant to brightness, contrast and rotation to some extent. Second, is the local feature vector quantization. We propose a N-nary coding quantization scheme instead of the binary coding method in the LBP operator. The N-nary coding method could preserve more structure information compared with traditional binary coding. Moreover, it still keeps the advantage that it does not require the cost of performing the nearest neighbor computation like the codebook quantization scheme. During the N-nary coding, we used histogram specification  to train the quantization thresholds.