Research And Application Of The Facial Feature Localization Algorithm

With the development of the Internet technology,biometrics technology has taken advantage of its stability and reliability,and more and more fields have begun to adopt this non-contact authentication method.Face recognition technology has become one of the hottest areas in recent years,for human face is the most easily captured human body part with the most obvious features.And the accurate positioning of facial feature points in face recognition technology will have a direct impact on the following work;therefore,the research on the localization of human face feature points has a profound significance.Based on the background,domestic and foreign scholars put forward various facial feature localization algorithms.Among these algorithms,the facial feature localization algorithms based on the cascaded regression and the shape search are most widely used.Though both the two approaches perform well on the existing date sets,they still have their own disadvantages.Currently,the main problems are the excessive dependence of the algorithm on the initial shape and the problems of occlusion and large deformations.The thesis mainly does the following researches.Firstly,for traditional cascading regression method is too dependent on the initial shape,we propose a method named face alignment based on classified shape searching.This method is based on the coarse-to-fine shape searching method.The traditional shape searching method requires every searching in the whole shape searching space,which involves a large amount of calculation.Solving this problem,we introduce a random forest classifier to classify the whole shape space into several sub-spaces.Images on the node of each tree in the forest are the same category.For each input image,we will find the closest shape searching sub-space,and the subsequent shape search will be performed in the shape search ing sub-space,which greatly reduces the amount of calculation and improves the experimental efficiency.Furthermore,this method ensures the accuracy and robustness of the algorithm.Secondly,in order to solve the problems of the SDM and DCNN methods,a face feature location method based on a cascaded stack self-coding network is proposed.This method proposes a new cascaded framework and introduces fou r basic cascaded stack self-encoding networks,including a global stack self-encoding network and three local stacking self-coding networks.The global stack self-encoding network directly gets a human face shape as an initial human face profile through a non-linear mapping,and the face we get is the input of the next local stack self-encoding network.For the local stack self-encoding network,we use shape-indexed feature as input,and the final face shape is close enough to the ground-truth face shape.Both methods are tested on the current challenging face data sets.The performance of the proposed algorithms is verified through the comparative analysis of experiments.