Research On Feature Extraction And Metric Learning Methods Towards Face Recognition

In recent years,with the development of computer vision and machine learning technology,great progresses have been made in the field of face recognition,demonstrating its value in many application scenarios.As two important technical parts in face recognition,face feature extraction and similarity comparison have a significant impact on the accuracy of face recognition.Deep learning based face feature extraction methods need a large number of manual annotated images,which is of high labeling cost.Besides,these methods generally perform not so well towards low-resolution face images.Metric learning is an important approach towards the similarity comparison.The existing metric learning methods generally suffer from the problems of insufficient discriminant power,complex optimization and computational failure.To address these problems,a series of studies on feature extraction and metric learning towards face recognition are conducted in this paper.To alleviate discriminative power insufficiency in current metric learning methods,this paper proposes an ensemble cascade metric learning method.This method enhances the discriminant power of distance metrics by cascade metric learning,in spirit of ensemble learning to prevent overfitting.In each cascade stage,the input features are randomly split into groups by dimensionality.Metric learning is executed in each group separately,and all learned distance metrics are finally integrated to achieving the good tradeoff between underfitting and overfitting.To address the problem of computational failure in some metric learning methods,this paper proposes a robust metric learning algorithm with approximate closed-form solutions,which has good scalability to both large and small data sets.Aiming at the problem that deep learning based face feature extraction methods need a large number of manual annotated training samples,this paper proposes a small dataset face recognition method based on feature disentangling.This paper first analyzes that disentangling facial features and noise is of great importance towards small dataset face recognition.Then the face feature disentangle network is designed.Four loss functions are employed as joint supervision to disentangle the face features and noise,which effectively improves the accuracy of the small dataset face recognition.Currently,most face recognition models are trained based on high-resolution images,and does not perform well on low-resolution images.To address this problem,this paper proposes a low-resolution face recognition method based on feature compensation.The proposed resolution sensitive branch is responsible for discriminate whether the input image is of low-resolution.The feature compensation branch is proposed to compensate the feature of low-resolution images and to align them with high-resolution feature in feature space via feature compensation.Making scarce harm to the recognition of high-resolution face images,this method improves the recognition accurate of low-resolution images significantly.