Research On Face Analysis Based On Multi-task Learning And System Implementation

Face analysis is a key technology in the areas of computer vision and deep learning.It has a wide application prospect in the fields of humancomputer interaction,biological information recognition,and emotion analysis.Face analysis research can be divided into three levels:first,locate the face region based on the input image,including face and background;second,recognize the attributes such as identity,age,gender,and race based on the entire face region;and third,analyze people’s expression and emotional state based on the specific area of the face.In recent years,the study of face analysis has made outstanding progress.However,there are still the following problems:(1)There is a training misalignment problem in the multi-task algorithm of face detection and face landmarks regression.There is a bias in the process of multi-task learning,which makes the convergence of a task poor,resulting in the decline of the overall performance of the multi-task network.(2)There is a problem of insufficient feature separation in face recognition and age estimation algorithm,which makes face recognition unable to effectively use age features to enhance the robustness of large age span sample recognition.(3)The problem of rapidly deformable samples in facial motion unit recognition leads to the low recognition rate of related states.Because of the existing problems,this paper focuses on the study of following aspects:1.Training aligned face detection and landmarks regression algorithm.In face detection and landmarks regression multi-task algorithms,the objective functions of the two tasks are difficult to converge synchronously.The positive samples selected in the training sample matching are biased towards the learning of the detection task and ignore the key point task,resulting in the non-convergence of the landmarks task.The nonconvergence of landmarks tasks affects the study of the overall framework,resulting in the reduction of face detection accuracy.This paper proposes a training-aligned face detection and landmarks regression algorithm(TAFDLR).This method aligns the multi-task training by aligning the scoring weight and learning accuracy of face detection and facial landmarks regression tasks in sample matching.Experimental results show that the performance of face detection in multi-task algorithms is better than that in single-task algorithms.Experiments are carried out on the WiderFace dataset.Compared with the baseline model without synchronous convergence,the average face detection accuracy is improved by 2.86%.Compared with the single-task model without landmarks task,the average accuracy is improved by 0.50%.Compared with the current mainstream face detection algorithms with the same amount of parameters and computation,the average accuracy of this method is improved by 0.68%on the test subset of small targets and fuzzy targets.2.Face recognition and age estimation based on feature separation.There are samples with a large age span in the face recognition and age estimation multi-task algorithms,which leads to a large distance in the extracted feature class,resulting in the decline of face recognition accuracy.Based on the assumption that the identification task is not related to the age estimation task,the main reason is the insufficient separation of age factors in the identity features extracted by the backbone network.A part separated face recognition,and age estimation algorithm(FRAEFS)is designed in this paper.This method separates identity features from age features through an attention mechanism to improve the robustness of identity recognition to age changes.The implementation of this method depends on multi-task learning,and the age estimation task is introduced to add constraints to the separated age features.The algorithm in this paper is tested on the dataset AgeDB-30,which contains samples with a large age span.Compared with the baseline work of a single task,it improves by 3.53%,which is close to the performance of mainstream algorithms in recent years.3.Encoder-decoder temporal network for facial action units recognition based on self-attention.The facial action units recognition algorithm realizes multiple binary classification tasks at the same time to recognize the states of multiple facial action units.There are rapidly deformation face samples in the task of facial action unit recognition.The existing feature extraction methods are difficult to describe the dynamic process of muscle rapid deformation,resulting in the decline of the ability of the network to distinguish different facial motion units.From capturing the discriminative information between frames and modeling the correlation within the time interval,this paper proposes facial action unit timing coding algorithm based on the selfattention mechanism(ETNet).Based on the self-attention mechanism,this method captures the characteristics of facial muscle changes when facial action units are activated by capturing the significant information between frames and modeling the temporal motion changes.The recognition tasks of multiple facial action units are realized by one network.Experiments on the DISFA dataset show that the proposed algorithm has advantages in recognizing facial motion units related to rapid deformation.The average F1 score of multiple facial action units is 1.5%higher than the existing methods.Through the above research,this paper constructs a multi-task face analysis algorithm,including face detection,landmarks regression,face recognition,age estimation and facial action units recognition.It improves the performance and efficiency of multi-task face analysis.And this paper designs a demonstrable multi-task face analysis system.