Research On Face Alignment Algorithm Based On Local Texture Models

With the increasing development of the artificial intelligence technology, people constantly refresh sense of cognition about the security, convenience and the trend of the identity authentication technology. Traditional identity authentication methods such as key, IC card and so on, exist shortcomings and many security risks, which doesn't meet people's pursuit of a reliable and convenient identity authentication. In recent years, the rapid development of personal identification technology based on the biological characteristics of human body, which dues to its wonderful safety and convenience. Comparison with other biometric identification methods based on fingerprint, palmprint, iris, retina, vein and bone, etc, face recognition deeply has the characteristics of non-aggression, anti-fake, convenient collection and so on. Face recognition system generally includes four steps:image capture, face detection, face alignment and identification. The goal of face alignment is to locate the feature points of the main organs such as the eyes, nose, mouth and face contour. Accurate location facial feature points is able to lay the cornerstone for accurate facial feature extraction and fast location facial feature points can effectively provide guarantee for real-time work of systems, which is more useful for face recognition work carried out smoothly. However in the complex environment, due to the influences of varied illumination, occlusion, posture, facial expression, image quality and other factors, the performance of face alignment algorithm greatly decreases, a robust and efficient method for face alignment is still a very challenging task.This paper focuses on discussion the face alignment methods based on local texture model. Firstly, elaborating the face detection based on Haar-like feature and AdaBoost algorithm. Nextly, descripting the face alignment based on active shape model in detail. Finally, stating the proposing face alignment algorithm based on random forest regression. These form a set of complete automatic facial feature point location system where the original image experiences from face detection to feature point location, the paper is mainly divided into four large pieces:Firstly, elaborating the research status and difficulties on the face alignment methods at home and abroad, and summarizing the general approaches and databases for face alignment work. Besides, summarizing the methods for face alignment performance assessment.Secondly, elaborating the implementation process of the face detection method based on the Haar-like features and AdaBoost algorithm, and explaining the extraction of Haar-like features, the construction of integral image, the construction and the cascading of classifiers in detail, and some experiments were performed in the MATLAB R2009A platform and Helen database, LFPW database and BioID database.Then, elaborating the implementation process of face alignment method based on the active shape model, and stating the aligning of the training sample shape, the construction of shape model, the extraction of gradient vector features and the construction of local models in detail, and some experiments were carried out in MATLAB R2009A platform and Helen database, LFPW database and BioID database.Finally, elaborating the implementation process of the face alignment method based on random forest regression proposed by us, and describing the extraction of pixel difference features, the construction and traversal of random forest regression models, the construction of global shape optimization model and the application of cascade structure in detail, and the proposed method has been tested in the MATLAB R2009A platform and Helen and LFPW database. The experimental results show that the face alignment method based on random forest regression is not only short time for feature points location but also high location accuracy, especially, the small training model is more valuable.