Unbiased Facial Age Estimation Based On Relational Reasoning

With the development of computer technology and artificial intelligence,the potential application of facial age estimation has attracted widespread attention such as advertising delivery,human-computer interaction and animation designing.Although existing face age estimation methods have excellent performance under constraint conditions,the generalization is limited for real world due to that the aging process is personalized for different aging images and different age classes have different collection difficulty.As a result,the biased data brings about the uncertain prediction,this dissertation aims at mining potential relationships from age data and control the latent influencing factors of aging based on relational reasoning.This dissertation focuses on the following researches:1.To solve the problem of heterogeneous age samples,a neighborhood-reasoning label distribution learning method is proposed.This method leverages the similarities between the age features to disentangle the latent semantic factors and build the neighborhood subgraphs.Then,exploiting the neighborhood relationships of subgraphs to estimate the corresponding label distributions.After that,the accurate facial age can be estimated using the expectation-maximization algorithm for mixed Gaussian distribution.Extensive experiments show that the proposed method is robust to heterogeneous face samples with different races,postures,expressions and illuminations in the real world.2.A progressive relational reasoning architecture is proposed for long-tailed age classification.T he method systematically calculates the intra-class variations,inter-class variations and class centers of the corresponding age classification and adaptively imposes various margins for each age classifier.The experimental results validate that the proposed method effectively balances the feature representation of each classification in the long-tailed data and improves the accuracy of age estimation.3.Based on the relational reasoning facial age estimation methodology and the principles of usability and robustness.A system prototype with the corresponding user interface including the visual analysis,training algorithm,test results function is designed and implemented.This prototype can effectively improve the generalization and practicability of the proposed method.