Face Key Point Detection And Face Attribute Analysis Based On Deep Learning

Facial landmark detection and face attribute analysis are two important research domains in computer vision.They aim at analyzing and understanding the rich informa-tion carried by the human face in digital images or videos and have been widely used in intelligent video surveillance,transportation and commercial face recognition system,etc.Facial landmark detection is the process of locating a pre-defined set of key-points for an input face image.By key-points,we mean some points around the contour of a hu-man face.This technology serves as the basis of a wide range of applications,including face parsing,facial expression analysis and face recognition.Face attribute analysis aims to classify each pre-defined face trait into two classes,i.e.present or absent.Accurately face attribute analysis can facilitate face identification by narrowing down the search space of face images in the database using face trait as the index.So efficient algorithms of facial landmark detection and face attribute analysis have crucial practical significance.Recently,deep learning models especially convolutional neural networks(CNNs)have achieved great success in various vision tasks such as image classification,object detection and semantic segmentation.Designing task-oriented deep learning model has been paid more and more attention to by both the academic and industrial communities.In this thesis,we propose different deep learning models for the facial landmark detection and face attribute analysis tasks respectively.Specifically,1.For the facial landmark detection task,existing related methods are reviewed firstly.To address the problems of existing methods based on deep learning in modeling the context of key-point and the structure among key-points,we propose a context module based on a context network and a structure module based on a tree model respectively.The context module can measure the context difference between the ground-truth and predicted landmarks,which is formulated as a part of the cost function of the model.While the structure module constructs a hierarchical tree to fit the distribution pattern of facial landmarks and then develop a loss function called structural loss to measure the deformation cost between the ground-truth and predicted hierarchical tree.Finally,we demonstrate the effectiveness of our method by the excellent experimental results on benchmark datasets.2.In the face attribute analysis task,a large body of research works emerges to cope with the task under unconstrained conditions.Face attribute analysis under uncon-strained conditions is challenging because of the drastic face appearance changes(caused by complicated factors like poses,occlusions and expressions).To address this problem,these works make great efforts to 1)design robust and discrimina-tive feature representations and 2)build effective attribute mapping functions.We focus on how to establish a unified end-to-end deep learning model,which is ca-pable of jointly learning discriminative features in conjunction with adaptively constructing attribute mapping functions.Extensive experiments on benchmarks demonstrate that the proposed approach achieves a better performance than the competing approaches.