Research On Face Detection And Face Alignment Based On Deep Learning

Face detection and face alignment are meaningful in computer vision that detect human faces and face landmarks in digital images.Face detection and face alignment are import procedures for face images analysis,and widely applied in face recognition and emotional analysis,etc.Deep learning has extraordinary ability for feature representation and robust in analyzing visual imagery.Applying convolutional neural network for feature extraction,which improves the detection rate of face detection and the accuracy of face alignment.In this paper,applied Deep Learning to enhance the accuracy of face detection and face alignment under unconstrained conditions.For face detection,applying deeper network structures can extract better features with the powerful functional approach.However,as deep convolutional neural networks being deeper and deeper,there are come some difficulty problems as gradient disappearance and gradient explosion which could have a bad impact on back propagation,preventing feature extraction and weaken face detection robustness under unconstraint conditions.On the other hand,since the human face alignment can be greatly influenced by imprecise face frames and complicated backgrounds,lights,gestures,expressions,occlusions,noises,etc.Mapping from face appearance to face shape became a kind of non-linear relation.Face alignment in specific video surveillance environment did not accurate enough and we have to create better model of increasing accuracy.Thus,this article focusses on developing efficient method for face detection and face alignment,analysis to the above problems,the corresponding perfection measures have been put forward in the thesis.The main innovations of this thesis are as follows:(1)In order to solve the problem of gradient disappearance and gradient explosion,we proposed two kinds of networks,which named Difference Networks and secondorder Difference Networks.Identity mapping with short-cut connections combined with original features and upper semantic information can avoid gradient disappearance and gradient explosion.Also,Difference Networks extracted more complicated features through short-cut connections with a subtraction operation.And,furthermore,the thesis considers Difference Networks in second-order way further improved features extracted in a difference method which was named as second-order Difference Networks.The experiment result of applying Difference Networks and second-order Networks show that this kind of structures avoid gradient disappearance and gradient explosion in deep convolutional neural networks and obtained more complicated network features which have better performance in feature extraction in CNN.Furthermore,it also enhanced the detection accurate rate in face detection,object classification and recognition.(2)For unconstrained conditions,consider difference between face appearances and shapes.Carefully analyzed the training dataset and test dataset,a targeted data augmentation with vary kinds of images was conducted,enrich diversity of samples in different scenarios,and improves the robustness of the trained models under unconstrained conditions.At the same time,we proposed a improved “coarse-to-fine” cascade shape regression network structure for face alignment.It includes: increasing the width of the network,added multi-scale feature fusion to enhance the ability of network feature learning in the global regression stage.It also proposed a novel loss function which including a parameter can be learnt at CNN,further improved regression capability of model trained.The experimental results show that the network with specific structure and regression loss function effectively reduces the error rate,improved the accuracy of the model in unconstrained environment,increased the rate of convergence either.