Multi-pose Face Image Processing Method

In this paper, face image processing methods utilized in multi-pose face recognition are presented on the basis of analyzing and comparing traditional methods in this field. Besides, the paper focuses on the standardization of face image, the normalization and synthesis of multi-pose face image from a single view. The main tasks are as followings:(1) The standardization of face imageBased on traditional integral projection method, two improved arithmetic are brought forth. One is the standardization of face image grounded on wavelet decomposition; the other is the standardization of face image grounded on skin-color model. The former combines traditional integral projection with wavelet decomposition. As a result, it has the same merit of simplicity, and the same demerit of sensitivity to pose as traditional integral projection. For that reason, it only adapts to front or slight rotating gray-face image. Despite the defects, the method can improve approximately 3 times at speed than traditional integral projection. The latter presents the standardization of face image based on skin-color model for face images screened under the natural surroundings because face images screened in that case can't meet the expectations of traditional integral projection method. It overcomes the demerit of sensitivity to pose, and has wider future in practical application, especially in multi-pose face recognition.(2) NormalizationThis section consists of gray normalization and size normalization. Histogram match method is used to realize gray normalization in this paper. Size normalization is realized by steps as followings: firstly, locating human eyes; secondly, rectifying the rotation in a plane; thirdly, cropping images to 1.7d×1.7d(Here, d is the distance between people's two eyes) taking the midpoint of the distance between the two eyes as base point and the distance between the two eyes as referent criterion; finally, zooming above images according to the quotient 1.7d/64(hypothesizing the size of standard images is 64×64). (3) The synthesis of multi-pose face image from a single viewFirst of all, the human face is represented with a dominant point set. Then, the variance of the dominant point set between different poses is fitted on the basis of a least square fitting with a polynomial function and a global morphing field is formed. Finally, multi-pose face images are synthesized by image warping from a single view based on the global morphing field. Since the rotation view of face image used in the whole research can't be obtained precisely, the only job we can do here is qualitative analysis, not quantitative analysis.