Research On Face Alignment Algorithm Using Supervised Descent Method Based On Improved HOG And Its Shape Optimization

As a critical step in face recognition process,face alignment accurately locates the facial keypoints position on the basis of face detection,and provides accuracy guarantee for face recognition,face tracking and attribute analysis.The face images are easily affected by factors such as head posture variety and illumination,which lead to inaccurate location of facial keypoints.Thus,face alignment is a popular research topic among scholars at present.Based on face alignment model using Supervised Descent Method(SDM),this thesis mainly focuses on how to solve the shortcomings of SDM algorithm,to reduce the error rate of facial feature points location and improve the stability of the algorithm.Main work of this thesis is as follows.(1)Aiming at the existing problems about SDM,such as inaccurate description of facial information by the extracted feature and the poor final alignment effect resulted by local optimum,an improved SDM(ISDM)algorithm based on improved Histogram of Gradient(HOG)feature and Social Spider Optimization(SSO)method is proposed.Our main contributions are twofold: Firstly,the existing feature extraction methods are discussed.An improved multi-scale HOG(IMHOG)feature extraction method is proposed to construct feature extraction window for each keypoint of the face.The size of the extracted feature window is determined by the alignment error in the iterative process.When the estimated shape is far from the real shape,the window size is set to be large,which can be estimated by utilizing the more available information;otherwise,the window size is set to be small,which can avoid calculation of redundant data.Secondly,to make the estimated shape close to the real shape and achieve global optimum,the popular SSO is applied to optimize the estimation result in iteration process globally.The experimental results have shown that the proposed ISDM face alignment algorithm based on IMHOG feature and SSO intelligent method is better than most previous algorithms on the common datasets,and meanwhile each part of improvement is verified to be effective.(2)As a regression model,the existing common problem is that the algorithm depends on the selection of the initial shape,especially under circumstances of head position posture rotation,exaggerated expression and partial occlusion.When face images are far from the average face shape,it is difficult to correct the deviation in the later regression learning.Thus,an optimized ISDM(Optimized-ISDM)algorithm based on initial shape selection is proposed.Firstly,the selective mode of initial shape in the original algorithm is improved.For each face,the average face of all samples is no longer selected as the initial shape,but the most similar facial shape in the initial shape candidate area will be chosen as the initial shape to train.At the same time,the local apparent texture information is only used in ISDM,and the global shape constraint is not considered in the regression function.Therefore,the local appearance and global shape information are combined to build the model.Redefining the alignment function is to make the regression function change according to the difference of the current shape.Experimental results have illustrated that the proposed Optimized-ISDM algorithm can improve the alignment accuracy of the whole algorithm,and be superior to the ISDM algorithm and most of the previous algorithms.