A Face Alignment Algorithm Based On Threshold Convolutional Network And Cascaded Regression

Facial landmark location is an important part of face recognition system,involving a number of fields of applied mathematics and computer vision.However,most algorithms are only designed for samples in small poses(below 45 degrees),and the commonly used face landmark location models assume that all the landmarks are visible,the invisible landmarks in large-pose samples have to be guessed.And the appearance varies more dramatically across large poses.Therefore,in this paper,we propose a respectively algorithm that can solve the large-pose face alignment based on the deep learning and cascade regression.The main contents are as follows:First of all,considering the large shape and appearance variation under extreme head poses and rich shape deformation,we propose three classification methods which are based on principal component analysis,pupillary distance and pose estimation respectively to divide the optimization space into multiple regions,and train the models separately in the sub-region based on the deep convolution neural network.The final predicted shape is estimated by the region model which the samples belong to.Experiments show that the sub-threshold convolutional network can make the large-pose samples be effectively learned as the same positive samples,so as to improve the overall location accuracy.Secondly,we propose a new initial shape selection method for cascaded regression model,which combine the k-means clustering with supervised classifier based on HOG feature.Each test is assigned an initial shape close to the global value.According to Standard SDM algorithm,a K-SDM algorithm is constructed,and the validity of this initial selection algorithm is illustrated through comparative experiments.Finally,the experimental results show that the mean error of K-SDM model in the 300-W databases is 6.24%,which is 1.26%lower than the standard SDM,which proves that the improved initial value algorithm can greatly improve the accuracy of cascaded regression to solve the large-pose landmark detection.In addition,the mean error of the sub-threshold-based cascaded deep model is 11.96%and 3.56%respectively on the highly challenging large-pose databases AFLW and 300W-LP,compared with the positive errors of 12.30%and 6.50%on the same scale,it can be proved that the threshold segmentation algorithm can make the large-pose samples have the same effective learning as the face samples and even have lower landmark errors,which is better than the existing ones method.