Digital Signal Processing Reference
In-Depth Information
and “retinex” model. Compensation method which is simple and efficient, however,
has difficulty with different lighting conditions and also, difficulty in achieving satis-
factory results in practical applications. Face modeling can get excellent recognition
rates. Otherwise, the algorithm is computationally intensive and is not very practical
for usage in real time face recognition systems. Accordingly, it is very difficult to
apply the approach in reality. Compared with the other two approaches, extracting
illumination invariant features is a more effective approach for face recognition under
various lighting conditions.
In order to obtain key facial features from face image under varying lighting, many
approaches apply multiscale analysis method to extract key facial features in coarse-
to-fine order. Xiaohua Xie et al. [5] suggest that illumination normalization should be
performed mainly on large-scale features of the face image rather than on the original
face image, proposing a method of normalizing both the Small- and Large-scale fea-
tures of the face image. Lu-Hung Chen et al. [6] utilizes the scale invariant property
of natural images to construct a Wiener filter approach to best separate the illumina-
tion-invariant features from an image. Goh et al. [7] proposes wavelet based illumina-
tion invariant preprocessing (WIIP) in face recognition. They decomposed a facial
image into low and high frequency components using discrete wavelet transform
(DWT) decomposition and set the illumination component as zero. Then both the
processed illumination component was used to perform inverse DWT. Haifeng Hu [8]
presented a discrete wavelet transform based illumination normalization approach,
which can obtain the multi-scale smooth images while preserving the illumination
discontinuities and can effectively reduce the halo artifacts in the normalized image.
Cao et al. [9] presented a NeighShrink-based denoising model (IIE). This model uses
neighboring wavelet coefficients to extract illumination invariant for face recognition.
Multiscale facial structure representation (MFSR) [10] attempts to normalize varying
illumination by modifying wavelet coefficients. Cheng et al. [11] proposed a method
employing NormalShrink filter in NSCT domain to extract illumination invariant
(NSNSCT).
In this paper, according to the NSCT multi-scale and multi-directional characteris-
tics, taking into account the neighborhood of subband coefficients within, a novel
illumination invariant extraction method was proposed to deal with the illumination
problem based on Nonsubsampled contourlet transform and NeighShtink denoise.
Experimental results on the Yale B, the CMU PIE face databases show that the pro-
posed method is robust and effective for face recognition with varying illumination
conditions.
2
Methodology
2.1
Retinex Illumination Model
According to Retinex theory, the image gray level can be assumed to be the product
[12]
Ixy Rxy Lxy
(, )
=
(, ) (, )
⋅
(1)
