Research On Facial Feature Analysis And Face Region Localization

Human face provides lots of information, thus draws a lot of attention on it. Inthis paper, we focus on facial feature analysis and face region localization.It is still a difficult problem to extract features from face images efficiently. Re-cent researches indicate much of the structural information of image may contain inthe high-order statistic among image pixels. A number of current face algorithmsuse face representations found by unsupervised statistical methods, such as PrincipleComponent Analysis (PCA), KPCA (Kernel PCA), Independent Component Analy-sis (ICA). Typically these methods find a set of basis images and represent faces as alinear combination of those images. These methods are very popular since the basisthey found could represent the high-order statistic well. In this paper, we use ICAand mutual information based method to extract local facial features from images. Inorder to analysis the properties of extracted features, we apply them to an age classifi-cation system which distinguishes adult faces from kid faces. The system consists ofthree parts: face detection, face alignment and normalization, age classification. Be-sides, we also use some other projecting method such as PCA and 2DPCA with LinearDiscriminant Analysis (LDA) to separate adult faces from kid faces for comparison.Experiment results shows that our method achieves 92.66% and 92.46% classificationaccuracy under 32×32 and 24×24 resolution respectively, and also improves clas-sification accuracy more than conventional 1D or 2D projecting method. Besides, weconstruct a system, which can detect faces in pictures and classify it to kid or adult,with trained model.When we want to extract and analyze facial features, we need face samples whichare cropped, and commonly normalized(aligned). In our work of local facial featureanalysis, we use the most popular face detection method to generate face samples, this approach works well on front faces in most case, but since this method is based onmachine learning, and has to confront some limitations. In resent years, researchersspent a lot of time on the process of visual recognition, and selective attention drawlots of attention for its active, selective and effective on processing data. In this pa-per, we combine selective attention mechanism with face region location, and aimsto make face region pop out in saliency map automatically. We modified incrementalcoding length model, which is based on features, to make it biased toward face region.We expand it into two-layer structure based on sparse coding theory, and use Gaus-sian pyramid to solve scale problems. Besides, in order to enhance saliency map andeliminate high contrast edges, we add region information and skin color information.Experiment results show that face region will pop out in saliency map by our model,and it biased toward face region much more than other saliency model. Moreover, ourmethod works fine on different scale and different pos face regions.