Research And Application Of The Facial Feature Localization Algorithm

Biometric identification technologies have a great prospect of development in that with the development of network technique and mobile communication technology,those fields such as the confidentiality of criminal investigation,the information security,and electronic transactions all are desperately in need of a more rapid and reliable information detection and identification technology.Because of the stability,the universality and the uniqueness of facial appearance,the contactless identity authentication and identification technology nowadays attracts more and more scholars to do researches in the area.Among the biometric technologies,the facial recognition technology is the most popular one.As to face recognition technology,facial feature localization is the most important step to implement the technology,because the accuracy of the facial identification in the following steps directly depends on the accuracy of the facial feature localization.Therefore the facial feature localization is of deep research value.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 the most widely used ones.Though the two approaches function pretty well on the existing date sets,they still have their own disadvantages.The approach based on cascaded regression depends on the setting of initialization when it regresses;while the operational efficiency of the shape search-based approach should be improved furthermore.Therefore the thesis proposes a new facial feature localization approach based on random forest to improve the operational efficiency of the above mentioned approaches.The thesis mainly does the following researches:Firstly,in order to avoid the low operational efficiency in traditional face search method,the thesis proposes a new facial feature localization algorithm combined by the random forest and the face search.The new algorithm functions as follows:at first,there would be a search space including various face shapes,which would be divided into smaller subspaces by using random forest classifier;then the present face shape sample would be distributed to the most related subspace according to random forest and face shape samples,and then in the related subspace the iterative face search begins immediately;what's more,in every search level there would be a related solution to restrain the present search shape;at last,the facial feature localization would be achieved by the cascade regression.The experiment shows that in the new method,the operational efficiency is much improved comparing with that in the traditional method.What's more,the accuracy and the robustness remain well in the challenging data set.Secondly,the thesis introduces three kinds of data sets that have been frequently employed to do experiments,and then creates a new data set based on the previous 300-W data sets.The new data set is extremely challenging in that it includes those images with big variation of head pose and severe occlusion conditions.Therefore,the new data set can test the robustness of the algorithm.In the meantime,considering of the different requirements towards the facial feature on different shape search stages,the thesis puts forward a new approach based on hybrid features,and combines the new method with the above proposed algorithm.The experiment proves that the operational efficiency of the algorithm has been greatly increased without losing much accuracy by employing the hybrid feature setting.