Facial Feature Points Auto Localization Based on Improved AAM Algorithm

开篇：润墨网以专业的文秘视角，为您筛选了一篇Facial Feature Points Auto Localization Based on Improved AAM Algorithm范文，如需获取更多写作素材，在线客服老师一对一协助。欢迎您的阅读与分享！

Abstract. In recent years， the facial features localization mainly uses active appearance model AAM. But when the face image effected by light， gesture， facial expression and image edge， angular position and other factors， the feature point fit is not efficient， feature point positioning accuracy declines. This paper poses an improved algorithm for AAM feature point location.Based on the reverse combination AAM algorithm， a shift-invariant wavelet transform method obtains low-frequency coefficients of face images， and image edge gradient features is enhanced by edge strength low-frequency coefficients.The gradient features is used as texture information.IMM face database simulation experiment by matlab language show that the improved algorithm is better than traditional AAM facial feature points position， the positioning accuracy improved by 7.69%.

Keywords： active appearance model；facial feature points；automatic positioning；fit

1. Introduction

The traditional AAM facial feature localization mainly composed of modeling and fitting.It uses gray scale characteristics to represent the image texture information.Apparent model is established by combining shape and texture information.Fitting is done by gradually narrowing the difference in texture surrounded by model texture and shape.This difference depends on the texture information of the training images sets.

But when the face image effected by light， gesture， facial expression and image edge， angular position and other factors， gray-scale characteristics of images will not be able to fully reflect the characteristics of the image texture and structure.The fitting accuracy greatly declines if only use gray features as the image texture features to create texture model.So， many researchers have made a lot of improvement.One of them is improving represent method of image texture.For example， Wolstenholme et al.applied wavelet transform to enhance image apparent ability and reduces the image dimensions which not only speed up the feature point positioning， but also guarantee the feature point location accuracy.Gao et al.， used Gabor filter to extract image texture information， and improved location performance.Another method is to increase edge or corner information by calculating edge intensity of each pixel and use the intensified information as texture.The AAM fitting precision is improved by normalized edge gradient information.Later Kittipanya - ngam adopted regional structure information to enhance AAM fitting ability.All the methods above improved the AAM fitting precision through rich image texture information or enhancing the image edge information.

The improved AAM methods separately processes texture extraction and edge characterization and are appropriate when the edge information is not affected by factors such as illumination， posture， facial expressions and image structure.But fitting is not efficient enough for complicated images.So， this paper puts forward an improved AAM algorithm.based on reverse combination AAM algorithm， low-frequency coefficients of face images is obtained by shift-invariant wavelet transform method， and image edge gradient features is enhanced by edge strength low-frequency coefficients.And the image edge gradient is used as image texture information.IMM face database simulation experiment show that the improved aam algorithm not only greatly improve the fitting accuracy but also speed up convergence speed and improved feature point localizing accuracy and localization speed.

2. Traditional reverse combination AAM algorithm

AAM uses the difference value between the generated model instance and the target image to guide the whole optimization process.AAM adds image texture information and has higher texture approximation accuracy when searches the target.The model not only considers the local characteristics information of the target object， but also adds the global texture information.It builds statistical model according to the face image shape and texture information from training sample set. The model accurately extract the image feature points.The basic idea of adopting this model to extract feature point is to use the model to generate synthetic image to approach the target image. By calculating the difference between synthetic image and target image to correct synthetic images in order to approach target image.The model not only extracts shape feature of the target object， but also realizes the texture expression of the target object.The fitting is an optimization process. It is widely used in facial reconstruction， expression recognition， target tracking and other fields.The traditional AMM fitting adopts gradient descent method to solve nonlinear optimization problems. This method has some disadvantages， such as laborious calculation and slow fitting speed.To solve this problem， Baker put forward AAM reverse combination algorithm based on Lucas - Kanade image alignment algorithm.

3. Improved AAM facial feature points localizing algorithm

The improved AAM facial feature points localizing algorithm is based on the reverse combination AAM algorithm.Firstly， it makes a shift-invariant wavelet transform to obtain low-Frequency coefficients of face images.Low frequency coefficients represent most energy of the image.It represent outline information of the original image， especially the information in the edge，corner point position and other key points.The high frequency coefficients represent the details of the original image.Then the edge gradient feature which will be used as image texture information is obtained by processing the edge strength low frequency coefficient.

4. Experiments and results analysis

Facial feature points localization experiment was implated in IMM face library with the improved AAM algorithm.There are 240 face images in IMM face library including forty people face.Each face is exposed in six different postures and light conditions.Each image was manually demarcated 58 feature points.Figure 1 shows parts of face image in the IMM face database.

First， AAM modeling was established according to the training set of face image and manual calibration feature points. Reverse combination algorithm and improved algorithm in this paper were conducted AAM fitting experimental respectively on the IMM face database.Figure 2 shows fitting results of the two algorithms.It can be seen from the diagram that the improved method in this paper increased the positioning accuracy of the face outside contour.This is due to the adding of edge intensity information which makes better convergence to face outside contour.

The positioning accuracy of the improved algorithm were measured by calculating the error between contour feature points and manual calibration points.A small error means high fitting accuracy and ideal match results.Figure 3 shows the average difference varied with the number of iterations.Table1 compares the two kinds of experiment result after 10 iterations.

It can be seen from figure 3， the improved AAM algorithm improved both fitting precision and speed.Because the low-frequency coefficients obtained from shift-invariant wavelet transform method provides more abundant texture information for AAM statistical model.It improves the ability of the AAM texture apparent.Also， low frequency coefficient after edge strengthen adds edge characteristic information to the image texture information.The edge information makes better convergence to facial outer contour and more accurate facial internal feature points location.Table 1 show that the improved method has lower average error after fitting iteration for many times， made a 7.69% increase.

It can be concluded from the experimental results that the improved AAM algorithm increases the AAM fitting accuracy by improving extraction method of texture feature.

Table 1 two AAM algorithm experiment results

after 10 iterations

5. Conclusion

Traditional AAM algorithm uses gray levels of pixels as pixel texture information， It can not fully reflect the characteristics and structure of texture characteristics of the facial image.In order to increase the facial feature points positioning accuracy speed， an improved AAM method was proposed.Based on the reverse combination AAM algorithm， low-frequency coefficients of facial images were obtained through a shift-invariant wavelet transform method， then image edge gradient feature is enhanced by processing edge strength low-frequency coefficients.The gradient feature is used as texture information. Experiments were conducted on IMM face database.The experimental environment：CPU is Intel Core2 Duo T5870. Frequency is 2.0 GHz， memory capacity is 2.0 GB， hard disk capacity is 500 G， Programming software is Matlab2008a.The experimental results show that the improved method gets more face image texture and edge intensity information， better convergence to the face outside contour， and more accurate feature points location.The improved algorithm is better than traditional AAM facial feature points， with positioning accuracy gained by 7.69%.

Acknowledgements

In 2013 Zhejiang Province Department of education research project（Y201330164），the 2013 Wen Zhou Science and Technology Bureau plan research projects （2013g0020）， the 2013 Zhejiang Dongfang Vocation and Technical College Key projects （DF201306） research results.

Reference

[1]Xinbo Gao， Ya Su， Xuelong Li， Dacheng Tao. Gabor texture in active appearance models[J]. Neurocomputing， 2009， 72：3174-3181.

[2]Kittipanya-ngam， T.F.Cootes. The effect of texture representations on AAM performance[C]. Proceedings of International Conference on Pattern Recognition，vol.2，2006 pp. 328-331

[3]C.Tomasi， T.Kanade. Detection and tracking of feature points[C]. In proceedings of sixth International Conference on Computer Vision Processings. Bombay， India， 1991：1-22

[4]S.Baker ， R.Gross and I.Matthews. Lucas-Kanade 20 years on： A unifying framework：Part 3. Technical Report CMU-RI-TR-03-35， Carnegie Mellon University Robotics Institute， 2003

[5]Fuzhen Huang， Jianbo Su， Yugeng Xi.Geometric Active Contours for Face Contour Extraction[J].Journal of Image and Graphics， 2008，13（01）：66-70.

[6]Zhigang Meng.Key Technology Research of Facial Nonrecognition[D].Southwest University of Science and Technology， 2007.