Feature Progression Model And Its Applications On Face Recognition

Face recognition has become a more and more popular human identification and recognition technology, including computer vision, image processing, pattern recognition, biological technology and other technical fields. Compared with other biometric technologies, face recognition is non-intrusive, and it can achieve better recognition results without disturbing people's normal behavior. In addition, face images can also provide information about a person's sex, age, race, etc., which is important in some applications. Therefore, face recognition has become one of the most promising biometric identification technologies, and has been widely used in security, tracking, identification and other fields. However, applying face recognition technology still has many difficulties, since there are many factors that may affect the accuracy of face recognition, such as the facial expression, and change with age,partial occlusion caused by glasses and hair, and feature covered or deformation caused by the light angle. Among them, face recognition across different poses is an important application in natural scene. Before remarkable progress in the study of this issue, face recognition application of high accuracy is limited to applications with input of frontal faces, such as customs, access control, face retrieval in identity database, etc. However, more other natural applications could not meet this requirement, like human recognition and tracking in video camera, face retrieval in video, images and on the internet. In addition, other face identification applications also require robustness to poses. In previous face recognition technology, face images need to be normalized with feature point alignment for effective target matching. The performance of algorithms could be seriously affected with mismatching of feature points, which is the main issue to be addressed by the pose-invariant face recognition.In this thesis, we focus on the drawback of relying on accurate pose estimation in prior pose-invariant technologies, and propose a feature progressing model to tackle the face recognition across different poses. In addition, we found a potential problem of high bias in face recognition experiments using cross-validation on databases with insufficient samples, when we tried to apply this model to age-invariant face recognition. We then carried out detailed analysis and discussion about this issue.The main contribution and innovations of this paper are listed as follows:1. In this paper, a new feature progressing model which describes the progressive change of human face due to specific factor is proposed, and is applied to pose-invariant face recognition. First, face images labeled with different poses are divided into different groups with equal pose gap. Second, we use multi-angle Active Appearance Model to detect and match the feature points, since human faces with different poses could not be uniformly aligned. Third, considering lamination variance caused by capturing faces in different poses, we adopted an illumination normalization method to obtain normalized human faces. The last is to extract features from the preprocessed face images, then input them into the feature progressing model and obtain the parameters. After the work described above, we conducted experiments on several public used cross-pose face image databases, and compared the performance of our method with the prior method in different scenarios.The results proved that the performance of our method is comparable to that of coupled-subspace method with probe pose known, while ours is better in scenario with probe pose unknown.2. We also attempted to apply the feature progressing model to age-invariant face recognition. Similar to pose-invariant face recognition, we first divided the age range into several groups, and labeled the face images with the group number. After face detection, the face region is aligned according to the position of several important feature points (eyes, nose). The next is to normalize the detected face with illumination normalization method and feature extraction with local feature descriptor.In the experiments, we found that Leave-one-out Cross-validation on small database(FG-NET) led to high false positive results, which needs further analysis. Another conclusion is that Partial Least Square does not apply to subspace construction on Feature Progressing Model for age-invariant face recognition.3. To discover the root cause of the high bias on recognition rate introduced by cross-validation on FG-NET, we conducted sufficient contrast experiments and provided mathematic analysis. The conclusion is that utilization of Principal Component Analysis and its variation as dimension reduction tools in regular face recognition experiments would bring high bias on the result of LOO-CV and K-fold-CV with large K. At last we offered available measures to avoid it.In summary, this thesis proposes a feature progressing model to maximize extracting discriminative identity components of facial feature in a special range of pose variation, and a pose-invariant face recognition method based on this model. We conducted a series of experiments in different scenarios to prove its robustness and effectiveness. In addition, we tried to apply this model to age-invariant face recognition. Although we didn't obtain ideal experimental results, we found a potential problem of high bias introduced by cross-validation in regular face recognition research. We provide a detailed analysis of such a problem and useful information on how to avoid such a problem.