Research On Machine Learning Methods And Applications For Facial Information Processing

The facial information plays an important role in identity recognition,head pose estimation,emotion analysis,age estimation,gender detection,etc.Compared with the fingerprint,iris and other biometric information,the face information collection process does not require physical contact,which makes the face information more suitable for application in investigation,security in public areas and so on.In recent years,with the rapid development of mobile Internet technology,the technology to deploy facial information processing algorithms to mobile devices has become a trend.Unfortunately,due to the restrictions on the computational resources,memory capacity,storage space,and power consumption,many algorithms can not be applied on the mobile platforms directly.To meet these problems,this paper focuses on the research of the face alignment,face recognition and facial feature extraction and try to improve the properties of these algorithm on accuracy,speed,and storage.The main contributions are summarized below.First,for face alignment,we introduce a new regression method which is called the affine transformation parameters regression(APR)to improve the training speed and the robustness of the supervised descend method(SDM).By combining the APR strategy and the traditional key points regression(KPR)strategy together,the hybrid approach produces a remarkable performance in term of accuracy,training time,prediction rate,and the model size in the face alignment task.Second,in the field of the deep learning,we propose a new training framework.There are two type of loss functions introduced in the new training framework,which are named the Label-Based Loss Function(LB-Loss)and the Feature-Based Loss Function(FB-Loss).By using different loss functions in different training stage,the training process converges quickly on any type of the convolution neural network with any depth.This makes it feasible that the researchers can customize the structure of deep network according to the application on mobile platforms and train it easily.Third,we combine the works above together and implement a face recognition framework based on the deep convolution neural network.With the storage volume of only5 MB,this implementation achieves the accuracy of 98% in the face verification task on the LFW dataset,and also achieves a good performance in practice,which can be applied on the mobile devices easily.At last,in the field of the feature extraction,we analyze the deficiencies of the average of synthetic exact filters method(ASEF),and propose a new feature extraction method.In experiments,the proposed method shows short training time,high prediction speed and remarkable prediction accuracy in the eye localization task.