Research On Algorithms For Facial Landmarks Detection

With the continuous development of computer vision and pattern recognition,re-search topics related to face recognition have received more and more attention.In the construction of face recognition system,accurate facial landmark alignment technology plays a very important role.The goal of facial landmark alignment is to accurately and quickly locate the important key points of a person's face,such as the corners of the mouth,nose,eyes,chin,and the like.So far,the facial landmark alignment have been extensively and deeply studied,and many practical research results have been obtained.However,due to the influence of pose,illumination,expression,and occlusion,it is still difficult to achieve a robust and efficient facial landmark alignment method.Based on the above research background,this paper systematically and deeply discusses the prob-lem of positioning of human face feature points.Then proposes some targeted facial landmark alignment algorithms and some satisfactory research results are obtained.The main research work and innovation of this article are as follows:1.After researching the problem of facial landmark alignment and the random forest construction method,an innovative local probability feature calculation method is designed.And then a facial landmark algorithm based on probability random forest combined with cascading regression method is proposed.Random forest is a decision-making method that has powerful generalization ability for data and powerful classi-fication ability for samples,and introducing it into our work can fully utilize its good performance.Cascading regression is to gradually approach the true value of the regres-sion goal with its stable and reliable performance.First,calculate the local probability features using the distribution properties of the training samples.Then,the calculated local probability features are used to train the regression model so that the prediction results more closely approximate the true location of the feature points.After that,the above steps are repeated a certain number of times according to the idea of cascading regression,so that the output of the training model gets closer and closer to the true value.Finally,combining different convergence models,the predicted position of the final landmark is given by the mean fusion method.A large number of experimental results show that the proposed algorithm has certain advantages over other classical algorithms and can overcome various changes in posture,lighting and expression.2.Based on the related research results of deep learning,a deep convolutional neu-ral network based on branch structure is proposed.And it is combined with the existing classical classification network+regression method to realize the positioning of facail feature points.In order to obtain accurate feature point positioning effect,in addition to the overall prediction of facial feature points,the algorithm divides the human face into two parts:the outer contour part and the inner feature part according to the distribution of human face feature points.In addition,in order to avoid the instability of a single model,multi-model fusion is also used in the final decision.Experimental results show the effectiveness of our algorithm on various face alignemnt datasets.