A Study On Head Pose Estimation And Human Age Estimation Method Based On The Face Image

Head pose estimation and human age estimation based on face image refer to enable the computer to automatically discriminate the pose and age label from a face image, which have become one of the most important topics in the field of computer vision research due to their significance to many real-world applications, such as human identity authentication, safety monitoring, human computer interaction and so on. However, the face image based automatic head pose estimation and human age estimation are still two very difficult and challenging research topics due to many image variations factors such as facial expression, illumination, scale, identity variabilities and so on, though many study works have been done by researchers.Generally speaking, there are two main steps in both the head pose estimation and human age estimation process. One is the extraction of discriminative face image features, the other one is the building of the recognition models. Therefore, based on the existing research work at home and abroad, in this article, an in-depth research into the2D image based head pose estimation and age estimation problems is conducted from the view of extracting a more discriminative features for the face image and recognizing the pose or the age by the machine learning methods. The main study contents and innovative work of this article are shown as follows:1) Propose a new face image representation method called Lie Algebrized Gaussians (LAG) feature. The first step of constructing LAG feature is to generate dense image patches and to extract shape, color and texture kernel descriptor features from image patches, which can capture the pose characteristics in appearance among different posed images or the aging patterns in appearance among different aged images. The global appearance representation of each face image is modeled by the local feature distribution of all its patches using the Universal Background Model (UBM) adapted Gaussian Mixture Models (GMM). Since it has been proved that the Gaussian probability density function (pdf) forms a Lie group, LAG is derived by mapping each Gaussian component of image-specific GMM to Lie algebra at the position of the corresponding UBM component, which has the capability to preserve the structure of the image-specific GMM in the original Lie group manifold. As a result, LAG is actually a representation approach encoding both the appearance information of the face image and the manifold structure information of the appearance in the feature space. Moreover, the extracted manifold structure of the feature space may be more likely to relate to the pose manifold of the pose image or the aging manifold of the age image. Motivated by the characteristic of the LAG feature, it is employed as the face image representation approach which is able to achieve better performance for the head pose estimation problem and the human age estimation problem.2) Propose a multi-scale LAG feature representation method for the face image. The multi-scale LAG is derived by dividing a face image into multiple sub-images in a spatial pyramid way and extracting LAG feature from each sub-image. The multi-scale LAG can capture more local information from the face image, which is helpful for head pose estima-tion. However, the dimensionality of the multi-scale LAG is high. In order to address this problem, the multi-scale LAG feature is optimized by the Principal Component Analysis (PCA) dimensionality reduction technique and then the head pose estimation is performed by the Nearest Centroid (NC) classifier. The Nuisance Attribute Projection (NAP) is further employed to enhance the discriminating power of multi-scale LAG for predicting the head pose label of a face image, which can use the pose label information of the face image and make the feature comparison more closely related to the pose.3) Study the Support Vector Machine (SVM) classifier for the head pose estimation problem based on the multi-scale LAG face image representation method. Considering that there are lots of pose irrelevant features in the face image and such noises are not eliminated in the multi-scale LAG feature for the face image, the Within-Class Covariance Normaliza-tion (WCCN) based SVM classifier is employed to perform the head pose estimation task in order to overcome the above defect. Compared to the general multi-class SVM classifier, the WCCN based SVM classifier is able to eliminate the similarity comparison of the pose irrel-evant features in the SVM classification process, thus it makes the head pose classification more accurate.4) For the human age estimation problem, after the analysis of the grading and ordinal characteristic of the face image, a new two-stage human age estimation approach which is in a coarse-to-fine fashion is also proposed in this article. This approach first obtains a rough age group for each face image by adjusting the coarse estimation value from a global SVR, then a local WCCN based SVM classifier is learned among the selected age classes in its age group to determine the final age of the input face image. This method avoids setting the age groups manually by looking for the age group for each face image automatically, which can obtain more precise age estimation during the local classification process.