Research Of Deep Bi-directional Ladder Network For Facial Attribute Recognition

With the explosive growth of face image data and increasingly complex practical application requirements in the big data era,facial attributes,the high-level image vision semantics,have attracted the attention of many researchers.This trend makes current computer vision algorithms focuses more on facial details and semantic meanings of face images,leading to the further development of massive related computer vision tasks.Facial attribute recognition is the basic task of facial attribute study,aiming to precisely identify the visual semantic information in given face images.Recently,the rapid development of deep learning has been driving the research on facial attribute recognition.However,the problems,such as insufficient usage of semantic information and poorly discriminative facial attribute features,have not been well addressed by existing methods.In light of this,this research proposes a deep learning based facial attribute recognition framework,with further proposed feature optimization schemes.The contributions of this research are:(1)Propose a Deep Bi-directional Ladder Network(DBLN)for facial attribute recognition.The hierarchical characteristics of the extracted features from the deep convolutional neural network(CNN)are considered to model the inherent intrinsic properties of facial attributes.That is,the local and global characteristics of the face attributes correspond to the features from the shallow and deep layers of the deep CNN,respectively.Experimental results show that DBLN can fully extract and discover the semantic information from different hierarchies of deep networks,leading to discriminative features that benefit for the facial attribute recognition.(2)Develop two facial attribute feature optimization schemes,resulting in Dual feature optimization based DBLN(Dual-DBLN)model for deep facial attribute recognition.On the one hand,with the special bi-directional ladder structure,this research design a dual residual attention mechanism based bi-directional feature optimization module for learning feature representations with meaningful semantic information;On the other hand,this research derives a local mutual information maximization constraint based single-directional feature optimization loss for learning more discriminative features.The extensive experimental results show that Dual-DBLN achieves state-of-the-art performance for facial attribute recognition.