Face Attribute Analysis Based On Multi-branch Network

Face analysis is an important research direction in Computer Vision.It has a wide range of applications in the fields of security monitoring,financial security,live entertainment and assisted driving.There are two main research methods for face multi-task analysis: one is a singletask learning method,and each task is designed with a dedicated model;the other is multitask learning,and a model is designed to solve two or three tasks at the same time.The single-task learning method superimposes multiple models when solving multi-tasks,which will increase storage and computation time;the number of multi-task learning tasks is relatively limited,each sub-task cannot converge at the same time,and the speed cannot meet real-time running requirements.This dissertation comprehensively considers the correlation between face attributes,and conducts a series of improvement studies based on existing face analysis methods,mainly including:1.The data is pre-processed,face detection is performed using an open face detector,and the face area is enlarged to expand.To address the difficult positioning of key points such as people's side faces and wearing glasses in the assisted driving scene,Mobile NetLandmark network is designed and compared with the traditional method LBF and the deep learning method DAN network in parallel.The average error and error rate both reach the optimal effect,and the running speed is 180 FPS.2.Propose a multi-branch face network structure,and simultaneously perform five tasks: face recognition,gender judgment,age estimation,appearance attribute judgment,and facial expression classification.The face recognition network is used as the backbone network.Based on the differences between the local and global face features required for the remaining four tasks,age estimation and gender judgment are classified as high-level task features,and appearance attribute judgment and facial expression classification are classified as low-level.Task features are derived from feature layers at different depths of the backbone network for training.3.Multi-task learning and shared feature parameters are used to train the proposed multi-branch face network.Finally,the performance of each task and the mainstream model is tested on a common test set.The test data sets(LFW,Age DB-30,CFP-FF,CFPFP)can achieve the current best results;the accuracy of gender judgment can reach 98.7 on the Celeb A and Faces of the World test sets % And 92.93%;the age estimation accuracy on the Adience test set reaches 60.1%,which is close to the current optimal model;the average accuracy rate of the appearance attribute judgment branch on the selected 8 attributes reaches 94.5%;the average accuracy rate of facial expression classification It is 71.3%,which is 65.9% higher than the baseline experimental results of the data set,and only 0.9% lower than the accuracy rate of the VEGAC model.