Design And Implementation Of Facial Feature Point Positioning Method Based On Deep Learning

Facial landmark localization,as the main basic technique in the field of facial image analysis,has important theoretical significance and application value.In particular,robust landmark localization plays a crucial role in face recognition under non-constrained conditions.Conventional deep learning methods mostly require a cascading architecture of a complex and cumbersome deep model,and the rough initial conditions may cause deviation of landmark localization,so that the localization effect under illumination,extreme posture,partial occlusion and exaggerated expressions is not good.In view of the predicament of the current landmark localization and the outstanding feature learning ability of deep learning,this paper focuses on the design and implementation of facial landmark localization method based on deep learning.The main work of this paper is as follows:(1)This paper expounds the research status of deep learning and facial landmark localization,analyzes and summarizes the current research priorities and hot issues in detail,classifies the algorithm according to the model of landmark localization respectively,and points out the advantages of various algorithms and the problems to be solved.(2)Aiming at the fact that the existing methods mostly focus on fewer landmark and the training process is complicated,a multi-landmarks localization method based on single convolution neural network is proposed.From the aspect of facial integrity,there is no need for cascading multiple layers of network or dividing faces into several sub regions.The facial multi-landmarks localization can be achieved using only a single deep convolutional network.This method can accurately extract global advanced features and directly predict the coordinates of multi-landmarks on the face.In addition,the method is robust to posture,illumination,expression,and severe occlusion in the non-constrained case.(3)Landmark localization is not an isolated issue and is susceptible to a number of different but closely related factors.In this paper,we propose a facial landmark localization method based on multi-task learning deep network,which is divided into two stages.First,by using a two-stage cascade of convolutional neural networks to generate facial candidate regions for a given test image,a multi-tasking deep network is used to fuse the features of the convolutional layer and the pooling layer in the middle layer and finally obtain the results of multiple tasks,including face detection box,gender and landmarks and their corresponding visibility information.Secondly,in order to improve the accuracy of face detection during testing,a two-stage cascaded neural network is proposed to generate facial candidate regions from the test image and remove a large number of non-face sub-windows to eliminate Complex background interference,thereby improving the overall performance of multi-tasking deep learning network.The study of this paper has certain robustness to interferences such as partial occlusion,extreme illumination and posture transformation,and has strong learning ability.It has been applied in many realistic environments such as Agricultural Bank and Nanjing Metro.