Face Localization Algorithm Based On Semantic Feature And Texture Feature

Facial landmark localization plays a crucial role in face recognition.As the fundamental and key part in face recognition system,the accuracy of localization directly affects the performance of face recognition.Although facial landmark localization has achieved initial success in recent years,it still faces some tough challenges including non-uniform illumination,large variations on pose and expression,partial occlusions and etc.Based on the study on SDM(Supervised Descent Method)algorithm and the shortage of off-the-shelf optimization solution algorithm of SDM,in order to improve the weighted way of the weighted least squares,we propose an improved optimization algorithm of SDM based on outlier detection and enhance the robustness of SDM.Based on what we know about facial landmark localization problem and algorithms,considering limits of single feature that lack of expressiveness,we propose a new approach to localize landmarks with abstract semantic features and concrete texture features fused:Choosing SIFT(Scale-Invariant Feature Transform)feature as abstract semantic feature,based on the proposed improved optimization algorithm of SDM,we achieve the fine localization;Choosing the texture feature of AAM(Active Appearance Model)as concrete texture feature,based on the Fast-SIC(Fast-Simultaneous Inverse Compositional algorithm)algorithm for fast AAM fitting,we achieve the fine tuning for facial landmark localization.The novelty is:The fusion of abstract feature and concrete feature is explicitly proposed.Exploiting the complementarity between features compensates the poor expressiveness of single feature;We propose joint localization with SDM and AAM.Utilizing the global perspective of SDM and the local adjustment of AAM,improves the local optimum of AAM and the coarse convergence accuracy of SDM respectively.To illustrate that the proposed localization algorithm can achieve high accuracy and robustness,we do the test on LFPW,Helen and BioID facial landmark fiducial databases and compare our results with several state-of-the art approaches'.Specially,the average eyes localization error of the proposed algorithm on LFPW achieves the minimum 2.02%,outperform the comparative methods.