Design And Implementation Of Facial Multi-attribute Feature Fusion Analysis System Based On Deep Learning

As the main research branch of face analysis,facial attribute analysis has important theoretical research significance and application value,especially for rapid multi-attribute recognition of unconstrained faces in natural environment.It has a very broad application prospects in field of public security monitoring and face retrieval Traditional deep learning methods mostly target specific single facial attribute,which not only require additional face image preprocessing,but also the single attribute model that separates the potential relationship between facial attributes.At the same time,the complexity of traditional model also limits the application scenario.In response to these problems,this thesis has carried out the following innovative works:A semi-automated image annotation method assisted by cloud services is proposed.Some face images of road traffic environments are annotated,which not only makes the research closer to the real world scene,but also provides additional training data for subsequent research.In order to solve the problem that facial attribute analysis relies on face detection and alignment,a joint model of face detection and attribute analysis is proposed.By using the fusion loss function,multi-class training data and training strategies such as OHEM,the robustness of the model in complex environments is improved.Finally,the existing algorithms are compared with on the FDDB,AFLW and Prima datasets to verify the superiority of the proposed model.To analysis the potential correlation between facial multi-attributes,the improvement of multi-task learning objectives and the realization of the network is taken.The multi-attribute feature fusion analysis model based on multi-task learning is proposed to realize the multi-attribute rapid analysis and prediction of unconstrained faces.The accuracy is better than most similar algorithms on CelebA and LFWA datasets,while at the same time achieving a significant increase in speed.For the road traffic scene,the driver monitoring system based on facial multi-attribute analysis is designed by applying the previous research results.Through the analysis and statistics of facial pose and eye state,monitoring of driver's driving and warning of fatigue driving are realized.The effectiveness of the system was verified by tests on the YawDD dataset.This thesis combines the existing research and makes a lot of improvements.The proposed model not only performs well in complex real-world environments such as low resolution,complex illumination,and partial occlusion,but also has extremely fast running speed and tiny model size.The research in this thesis covers facial area detection,facial multi-attribute analysis and application,and runs through the complete research and application process,which has a positive effect on the application of facial multi-attribute analysis research.