Research On Human Facial Age Estimation With Underlying Correlation Exploitation

Human facial age estimation(AE)becomes an important research topic and has attracted a great deal of attention due to its wide applications in recommendation system,security mon-itoring and other scenarios.On the one hand,although plenty of related methods have been proposed,most of them pay attention to holding on the inherent characteristics of facail age at-tribute or exploiting the specific correlations between age attribute and other facial attributes.In contrast,few of them attempt to explore the potential relationships within facial age attribute.On the other hand,due to the incompleteness of annotated age labels in public human facial aging datasets,the generalization ability of the models trained by these datasets is seriously in-sufficient.To this end,with a perspective on correlation learning,we attempt to explore these relationships and embed them into AE models to impove the performance of AE.To sum up,the main contributions of this dissertation are summarized as follows:1)Exploiting the joint relationships between feature representations and age labels.In this dissertation,we construct two kinds of relationships self-learning models derived by Least Square Regression(LSR),so-called AELR and AEFR,respectively.AEFR can automatically exploit the underlying relationships inter-/intra- feature representations,which is rarely involved in the existing works.To avoid that the intrinsic structure between CA is destroyed indirectly,the relationships inter-/intra- age are exploited by AELR,automatically.And then,in order to improve the performance of AE,we propose a joint relationships self-learning model,namely AEJR,to explore and exploit the joint relationships between feature representations and age labels.Due to the influence of the gender attribute,we extend our proposed relationships self-learning models to gender-aware counterparts.Extensive experiments are conducted to testify that feature representations or age labels are related,and more importantly that the proposed methods modelled incorporated with these relationships can significantly improve the accuracy of AE.2)Exploiting the intra-class correlations and input-output relationships.To analyse the limitations of LSR and its variants,it is obvious that the AE models based on least squares loss function suffer from poor discriminating ability,which ignore both the intra-class struc-ture correlations and the input-output relationships.To address this drawback,we construct a structure-exploiting discriminative ordinal multi-output regression model with the correlation exploitation via a low-rank structure matrix and the metric exploration,namely SEDOMOR.The structure matrix depicts the input-output relationships,and the metric characterizes the intra-class correlations by dragging the regression margin automatically.In addition,to further enhance its distinguishing ability,we extend the SEDOMOR to its non-linear counterparts with kernel functions and in deep architectures for AE problem with highly non-linear data.Finally,extensive experiments testify the effectiveness and superiority of the proposed methods.3)Exploiting the cross-database structure correlations between various datasets.With the comprehensive comparison of the label information in multiple human facial datasets,it is obvious that the scope of the annotated age labels of these datasets are not complete.Thus,the AE models obtained from these datasets can only depict partial facial aging patterns,which leads to the serious limitation of its generalization ability.To overcome these issues,we propose a structure-preserved ordinal domain adaptation model,namely SPODA.In particular,a domain representation matrix and a manifold learning paradigm exploit the structure relationships in the aspect of cross-database and intra-class,respectively.Moreover,to boost the generalization of cross-database age estimation,the metric is introduced to drag the regression margin automat-ically.In terms of AE,extensive experiments indicate that the SPODA is more effective and advanced than the existing domain adaptation models.