| Face landmark detection known as face alignment is one of the classic problems in the field of computer vision.The goal of face landmark detection is to locate a set of pre-defined face key points,such as eye corners,mouth corner points.The current face landmark detection task only results in small positioning errors under limited conditions,meeting the needs of specific scenarios.However,with the development of scenarios,face images in different scenarios are obtained under various non-limiting scenarios.The images in these scenes vary from illumination,background,face pose,image quality,etc.which makes the landmark detection problem of the face difficult to solve well.With the development of convolutional neural networks,the existing face landmark detection network model design gets more complex,but the results are of them are not distinct from each other and the model operation efficiency is low,which results in the analysis and evaluation of the model getting more difficult.In addition,the existing regression-based face key point detection model does not take the integrity of the face landmark into account,which makes the model output results lack stability and be prone to outliers.This thesis focuses on the detection of face landmark under unconstrained conditions,ac-cording to this,we propose a streamlined face landmark detection network introducing a global constraint mechanism.The main contributions are as follows:1.The existing landmark detection network gets increasingly complicated.We adopt a new streamlined key point detection network structure,which has greatly improved its efficiency.The entire thin network is made up of a feature extraction module and an upsampling module consisting of several deconvolution layers.At the same time,Resnet's residual unit in the feature extraction network structure is replaced for better extracting image features.2.The existing network structure does not take the integrity of the face into account.This thesis adds global constraints on all key points of the face in the network structure to reduce the outliers in the prediction results.The experimental results on the 300-W dataset show that the detection performance of the reduced network structure with global constraints exceeds the current typical deep neural network detection model. |
