Research On Deep Learning Method For Face Alignment

Face alignment,or called face landmark detection is a popular research issue in Computer Vision,it is widely used in many visual tasks,such as face recognition,facial expression analysis,3D face modeling and age estimation,etc.And precise facial landmark locations play an important role in these tasks.Face alignment aims at automatically and precisely locating facial landmark points,such as eyebrow,nose,eye,mouth and cheek contours.At present,several face alignment methods based on cascaded shape regression model have achieved great achievements which are very close to labels in manually marked manners,however,face alignment is still facing great challenges due to the face images captured in unconstrained environment,which are easily affected by illusion,pose,expression changes and occlusion.Over the past few years,with the utilization of deep learning method in face alignment,the accuracy of locating facial landmarks has been greatly improved.In this dissertation,we propose two more robust approaches for face alignment,which is based on cascaded shape regression model and combine with the auto-encoder network and convolutional neural network respectively.The main research work and innovations of this dissertation are as follows:(1)By analyzing the cascaded shape regression model,we found that a good shape initialization strategy is very important for the subsequent regression process.Therefore,we divide the face alignment task into two stages,which are used to generate initial face shapes and then do fine-tuning in a coarse-to-fine manner to predict facial landmark points.The face alignment algorithms proposed in this dissertation also follow this framework.(2)In this dissertation,we present a stack auto encoder network model for face alignment.Through the research on difficulties of landmarks regression in different facial components,we factorize the global objective function as five local cost functions in the stage of face shape readjustment.We also analyze the shortcomings of the SIFT features,extract the local pixel features and improve results by fusing them together.At last,several stacked auto encoder network groups gradually update the face shape in cascaded manners.(3)In this dissertation,we propose a face alignment method based on two-stream convolutional neural network,instead of using a stack auto encoder network to predict initial face shape as previous method does,we take convolutional neural network to generate a better shape.In the stage of face shape readjustment,two convolutional neural networks with the same structure are arranged in parallel,which are used to extract shape features and image features respectively.Then these two features are combined by calculating the matrix outer product to obtain the fusion feature representation before fully connected layers.At last,several networks with above structure gradually update the face shape in cascaded manners.(4)In this dissertation,an intelligent neural network training strategy is proposed.By inputting the training set in batches,the actual test set environment is maximally simulated,the degree of overfitting is effectively reduced,and the generalization of the model is increased.Results of experiments on multiple datasets and comparison with other state-of-the-art methods demonstrate that our proposed method make face alignment more accurate and robust.