Research On Sample Representation And Cross-Domain Learning Method For Soft Biometric Estimation Task

Soft biometrics refers to the measurable,non-unique,and imprecise individual characteristics of the human body externally or internally,such as age,facial attractiveness,height,body shape,etc.These features are generally used in fields such as individual identification,behavior analysis,and security detection.Compared with traditional biometric features such as face,fingerprint,and iris,soft biometrics have the advantages of high privacy and wide applicability.Enhanced recognition,verification,and description of human subjects has thus received increasing attention in the field of deep learning.However,deep learning require sufficient and high-quality data and labels,while the task of soft biometrics estimation faces more serious problems such as data privacy,data permissions,and biological protectionism,and it is difficult to obtain large-scale labeled data to support the optimization of neural networks.Meanwhile,soft biometrics estimation are not simple classification or regression problems,which are difficult to solve using only classification or regression methods.Therefore,this thesis starts from the data shortage problem,combines the characteristics of soft biometric estimation and the advantages of methods in various fields,and improves in representation learning and estimation methods.Specifically,the main contributions of this thesis are as follows:(1)Aiming at the problem of insufficient training data,this thesis proposes a reconstructed feature estimation method based on independence and consistency losses.Under the constraints of independence and consistency losses,and combined with feature separation and recombination network,the features related to the estimation task are separated and then recombined with different faces to generate a large number of new feature samples for regularized training,the problem of insufficient samples is effectively alleviated.At the same time,soft biometrics estimation is neither a simple classification problem nor a complete regression problem.For the estimated features that are not constrained by the special attributes of faces,the method of combining label distribution and ordered regression is used to optimize.The mean absolute error results on CACD,Mega Age-Asian and SCUT-FBP5500 dataset are reduced by 0.24,0.24,0.031,respectively.(2)In order to promote the reconstruction-based representation learning method to a wider application,an adaptive network regularization based on attention is proposed.Through attentionguided feature reconstruction and regularized training of the network,the potential of the neural network can be fully tapped and the robustness of the convolutional neural network can be greatly improved.The lightweight feature reconstruction module can be plugged into different models of different tasks to further improve the prediction accuracy of image classification and soft biometrics estimation tasks without adding too much additional resources.(3)By analyzing the characteristics of the soft biometrics estimation task,this thesis propose an ordered verification method under the disordered estimation task.By constructing meta-tasks on soft biometric estimation,and a large number of disordered meta-tasks are trained to simulate the process of evaluating query image label from a set of examples.Query and estimation are carried out by combining ordered verifier and ordered regression classifier.In addition,based on the CACD,Mega Age-Asian,and SCUT-FBP5500 datasets,the proposed meta-estimation has good cross-domain and cross-task learning performance between different domains and tasks,providing methodological support for soft biometrics estimation tasks that lack training samples.