Research On Face Detection And Attributes Recognition Based On Deep Learning

With the development of deep learning and artificial intelligence,face detection has become the basis of major application algorithms.How to detect blurred,small-sized faces in complex application scenarios quickly and accurately is the focus of face detection in recent years.And the recognition of human face attributes can make humancomputer interaction have more and better experience.The purpose of this paper is to achieve the balance of accuracy and speed in the research of face detection and attribute recognition.After studying the progress of face detection and attribute recognition,this paper first proposes two fast methods of face detection.One is based on the single stage headless face detector which is called SSH algorithm,which adjusts and improves the anchor mechanism,adjusts the whole network according to the strategy of receptive field matching,improves the feature pyramid network by using dilated convolution,and finally improves the feature pyramid network.Through deep separable convolution,the parameters and running speed of the whole network are greatly reduced.Finally,our algorithm has reached m AP 0.956 on easy set,m AP 0.933 on medium set and m AP0.825 on hard set for Wider Face dataset.With less loss in precision,the fastest speed can be achieved to 16 ms per frame.The other one is based on MTCNN algorithm by adjusting the parameters and convolution structure,the detection speed is continuously improved.It is trained with Io U continuous upgrading strategy,which makes the MTCNN algorithm faster and achieves the availability in simple scenarios.In addition,on the basis of the study of face attribute recognition,multi-task facial expression and gender classification is carried out.In the process of age recognition,facial expression and gender features are fused with age features.PCA feature dimensionality reduction is made for the fused features.Finally,the complexity of training and testing is reduced by using extreme learning machine,the final model performs well.