Cross-Age Face Verification

With the demand for personal security and human-computer interaction in recent years,image-based face verification attracts many scholars and research institutions to engage in this research field and has remarkable advances achieved in both academia and industry.However,image-based face verification remains a challenging task due to the large number of variations caused by illumination,pose,age gaps and so on.Facial aging is a complex and continuous process,and it affects many variations.Thus,cross-age face verification is mainly faced with the following challenges:each individual may have different characteristics in different age stages;different individuals may share similar characteristics in the same age stage;cross-age face images contain different expression,and the differences of facial expression also affect the performance of face verification.In this thesis,three cross-age and cross-expression face verification models are proposed to solve above problems,the main contributions are summarized as follows.Firstly,we propose an ensemble face pairs distance metric learning for cross-age face verification.Considering the differences of facial variations for different individuals,it is difficult to learn a uniform aging pattern.Furthermore,the differences of face pairs with different age gaps are different.In this thesis,an ensemble method combined with face pairs distance metric learning is proposed to effectively reduce the influence of complexity caused by age gaps and improves verification accuracy.A series of experiments are conducted and the results show that the new method can significantly improve the performance of face verification.Secondly,an ensemble multi-scaled multi-layer model to represent face pairs for cross-age face verification is presented.Feature learning for representation of face pairs is crucial for face verification.To extract sufficient powerful high-level discriminative features,we put forward a new model by using multi-scale scanning and multi-layer feature extraction to capture high-level discriminative features.In addition,the existing discriminative face verification approaches usually take the difference between each pair of faces as the input data,which equally treats all pairwise differences but ignoring the variances among pairwise differences.To solve this problem,we take the original information of face pairs as the input and automatically consider the variances among face pairs in the subsequent learning stages.A series of experiments on FG-NET?CACD and CACD-VS datasets are conducted to demonstrate that proposed model can effectively reduce the influence of age gaps and outperform the existing methods.Finally,we designed an independent component learning model for cross-expression face verification.Face images contain different facial expressions,and the variation of different expressions also affects the performance of face verification.To reduce the impact of facial expression,we learn identity sub-space and expression sub-space independently and identify features of original face images are represented in identify sub-space.A series of comparative experiments on JAFFE dataset are conducted to demonstrate that the new model can effectively learn expression-invariant identify features and significantly reduce the influence of expression on face verification.