Face Landmark Detection In AR Application

As an important biometric feature,facial information has been widely used in academia and industry.In recent years,on the basis of higher theoretical level in the field of computer vision,hardware facility has also developed rapidly,which makes the speed and accuracy of face detection and more refined face landmark detection also rapidly increase.The face-based augmented reality application is based on the registered facial information,and uses the related technology to superimpose the virtual effect on the facial information under the real scene.The landmark detection is an important research basis in the fitting process.There are many scenes such as makeup in the application.At present,the positioning accuracy,efficiency and fitting effect of the landmark detection algorithm still need to be improved.Whether it is the try-on effect of the e-commerce sales platform or the makeup effect in the beauty camera,it is necessary to gradually improve accuracy from the constraint condition to the unconstrained transition.Therefore,this paper analyzes and studies the facial augmented reality application scenes about the characteristics of head posture and large facial expression changes in the complex background,and achieves the effect of the scene of “film and television drama substitute”.Based on a detailed analysis of the characteristics of this scenario,the main work of this paper is the following two aspects:1.A new method based on cascaded regression framework,CPR-LADF algorithm,is proposed to solve the problem that the output of face feature detection algorithm is greatly affected by head pose changes and complex background.And it is improves the speed of the algorithm in a way.In the process of extracting features,this paper adopts a variant feature of pixel difference feature,which is more stable than the pixel difference feature.Compared with other cascading when the head pose and face expression change greatly.The FPS3000 and other algorithms in the regression framework have better output results.In the process of training the random forest model,this paper takes 3-6 feature points as the vertices to form the local area,randomly selects the reference points in the local area to find the mean value,and then averages the values.As a feature of splitting,we create a random forest for each feature point,output the face shape variables of each stage through multiple iterations,gradually optimize the random forest model,and then use global linear regression and multi-model fusion to obtain the output results,and finally accurately locate Face feature points.2.Realized the AR effect of the "film and television drama substitute" application sceneThrough the improvement of the above algorithm,the detection effect of face feature points under unconstrained conditions is enhanced.The AR “face changing” effect is tested under the scene of "film and television drama substitute" with complex background characteristics.The research results can achieve real and false results.The effect of integration,which has opened up the solution to the existing problems in the film and television industry to a certain extent,such as the excessive appearance of stars,the unreasonable distribution of production costs and so on.At the same time,it can also bring better viewing experience to users,and has greater in-depth research and application value.Finally,the experimental scheme is given and analyzed.The improved face feature point detection algorithm has a certain degree of improvement in algorithm robustness and the speed of detection.There is also a certain promotion for unconstrained augmented reality applications based on facial information.