Research On Facial Landmark Detection Methods Based On Deep Learning

Facial landmark detection is a classic task in the field of computer vision,which has great academic research value and commercial application prospect.In recent years,the rise of deep learning technology has accelerated its research and development process.A large number of facial landmark detection algorithms based on deep learning have been proposed,among which two kinds of methods based on coordinate regression and heatmap regression are favored because of their superior performance.The coordinate regression method uses the image information to directly predict the coordinate value of the facial landmarks,while the heatmap regression method transforms the facial landmark detection problem into the heatmap estimation problem for solution.Generally speaking,the detection performance of heatmap regression method is better,but the time required to execute the algorithm is relatively long,while the detection performance of coordinate regression method is lower,but the time required to execute the algorithm is relatively short and easier to control.In this paper,the algorithm of facial landmark detection based on heatmap regression is well studied.Firstly,facial landmarks are detected using the classic stacked hourglass network.For the original hourglass network,the encoder's feature extraction ability is relatively weak,and the decoder does not make full use of the complementarity between features of different scales.Therefore,an improved stacked hourglass network is proposed by using a stronger and more general feature extractor to replace the original encoder and embedding a multi-scale feature fusion module in the decoder,which significantly improves the network's feature expression ability and achieves better facial landmark detection results.However,due to its non-differentiable post-processing operation,the output result of the above algorithm is discrete value,which has quantization error.There is still a certain deviation between the detected facial landmark's coordinate and the actual position.In view of this,this paper further proposes a two-stage facial landmark detection algorithm based on offset learning.Considering that the coordinate regression network has the characteristics of continuous output value,this paper uses the coordinate regression network to fit the offset between the prediction result of the heatmap regression network and the corresponding real label,and then integrates the coordinate regression network and the heatmap regression network together to further improve the performance of facial landmark detection.Moreover,a real-time facial landmark detection system in video is developed in this paper.The facial landmark detection algorithm is implemented using the lightweight coordinate regression network,which can realize real-time detection on the computer CPU.In view of whether the number of faces in the video will change,this paper presents two design schemes for the scene with the number of faces changing and for the scene with the fixed number of faces.The former adopts the strategy of detecting faces and facial landmarks frame by frame,while the latter only detects faces at the start frame of the video,and the face bounding boxes of each subsequent frame is calculated from the facial landmark's detection results of the previous frame and then expanded by a certain proportion,which avoids the process of detecting faces frame by frame,and further saves the processing time of the system.