Multi-pose face feature point detection method based on cascade regression

1. A multi-pose face feature point detection method based on cascade regression, comprising the following steps of:

(1) extracting pose index features and establishing corresponding optimal weak regressors: using a clustering algorithm to cluster face feature points to acquire feature point categories with adjacent positions; extracting pose index features under corresponding poses according to the feature point categories; and inputting the pose index features into a cascade regression algorithm, and training the pose index features to acquire the corresponding optimal weak regressors under different face poses; and

(2) performing initialization and detection on face feature points under multi-pose changes: performing corresponding initialization according to different face pose orientations; using an SIFT feature of a face image as an input feature for face orientation estimation; acquiring an orientation of an input face image according to a random forest face orientation decision tree; using a feature point mean value of a face training sample under the orientation as an initial value of the input face image feature point; and extracting the pose index feature of the face image and inputting the pose index feature into the optimal weak regressor to acquire a distribution residual to update the current feature point distribution, and complete the face feature point detection.

2. The method according to claim 1 , wherein in the step (1), the poses are divided into different training pose subsets according to the face orientation; in each pose subset, a k-means clustering algorithm is used to cluster the same category of feature points in the internal feature points of the face, and a mean value of a clustering center is continuously updated until a sum of square errors D=[*]Σ i=1 k Σ p⊂C i |p−m i | 2 convergence, wherein C is a clustered cluster, k is a number of clusters selected according to different pose subsets, i is a face feature point, and m i is a mean value of an i th cluster; and finally, the internal feature points of the face are clustered into five categories: left eye, right eye, nose, mouth, and chin.

3. The method according to claim 1 , wherein in the step (1), the process of extracting the pose index features under different poses comprises: in training subsets with front face orientation, randomly selecting three types of candidate feature points from the five categories of left eye, right eye, nose, mouth and chin; in training subsets with left side face and left face orientations, randomly selecting three types of candidate feature points from the four categories of right eye, nose, mouth and chin; similarly, in training subsets with right side face and right face orientations, randomly selecting three types of candidate feature points from the four categories of left eye, nose, mouth and chin; after selecting corresponding feature point sets, acquiring a mean value feature of a face triangular region composed of the feature point sets.

4. The method according to claim 3 , wherein the step of acquiring the mean value feature of the face triangular region comprises:

step a: acquiring local triangular region template sets: after selecting corresponding three types of candidate feature points in face training subsets with different poses, randomly selecting a feature point in each type as a vertex, and finally forming a closed triangular region by the three vertices, wherein the regions form N local region template sets L={L n ,n∈[1,N]};

step b: mapping reference points: mapping all templates in the set L to samples, randomly selecting a reference point Q, and determining a feature point distance by recording a distance d between the reference points and a nearest vertex Q′ in the local region to complete mapping the reference points; and

step c: solving the mean value feature of the local triangular region: after completely mapping the reference points, calculating the feature points, and acquiring a mean value

B

=

∑

j

=

1

M

⁢

⁢

I

⁡

(

Q

j

)

M

of M reference points in each closed region, wherein I(Q j ) is a pixel value of a point Q j in the image; and differentiating N local region mean values in the set L in pairwise to acquire an N 2 -dimensional local region mean value feature.

5. The method according to claim 1 , wherein in the step (2), in the face orientation estimation, the SIFT feature of the input face image is extracted as a classification feature, and input into the trained random forest decision tree, that is, a probability of each category is accumulated, and output as a category with a largest sum of predicted probabilities in all the trees, and categories with different face orientation are acquired.