【作者】 王峰；

【导师】 潘纲； 王跃明；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2017， 硕士

【摘要】 人脸对齐是计算机视觉中的经典问题之一,其目的是自动计算出给定人脸图像中的面部关键点坐标。精确的人脸关键点定位结果对许多视觉任务具有重要意义,如人脸识别、3D人脸重建、人脸表情分析、人脸姿态估计等。随着相关技术的发展,目前的人脸对齐方法在受控条件下可以达到较低的定位误差。然而,许多人脸相关应用的输入是在自然条件下获取的,由于存在光照、背景、人脸姿态、图像质量等多种干扰因素,人脸对齐问题依然非常具有挑战性。本文主要关注非受限条件下的人脸对齐问题,主要贡献点如下:(1)本文通过实验分析发现,合理的初始值可以使级联回归模型的定位误差率大幅下降。基于该发现,本文提出了由粗到精的两阶段人脸对齐算法框架,将人脸对齐分成粗定位和精定位两个子问题,且每个问题应该使用专用方法解决。(2)针对粗定位问题,本文设计并实现了一种基于深度卷积神经网络的模型,该模型以整张人脸为输入,直接预测所有人脸关键点的位置坐标。在300-W测试集上的结果表明,该模型能有效降低人脸关键点定位失败率。(3)针对精定位问题,本文提出了一种基于参数共享的级联回归模型,该模型中的每个回归步骤均使用相同的参数。与粗定位模型结合后,在300-W测试集上误差率降低到了 state-of-the-art。此外,本文还指出可以将单个回归模型作为梯度预测模型,通过结合梯度下降算法中的技巧,实验表明关键点定位误差率还可以获得进一步下降。更多还原

【Abstract】 Face alignment is one of the most classic problems in computer vision,which aims at localizing landmark automatically.Accurately locating facial landmarks is very meaningful for many visual tasks,such as face recognition,3D face reconstruction,facial expression analysis,face pose estimation,etc.However,the input images for many face-related applications are captured in unconstrained environments,and face alignment in such cases is still very challenging due to the large variations in background,illumination,pose and image quality.This paper mainly focuses on the problem of face alignment under unconstrained conditions.The main contributions are as follows:1)This paper presents the experimental results that reasonable initial locations of landmarks can lead to a significant reduction in localization error rate for cascaded regression models.Based on the findings,this paper proposed a coarse-to-precise face alignment algorithm framework,which divides the face alignment into two sub-problems:coarse localization and precise localization,and it’s suggested to design dedicated methods for each sub-problem.2)As for the coarse localization problem,we have designed and implemented a model based on deep convolutional neural networks,which take the whole face image as input and predict the coordinates of all facial landmarks.The results on 300-W test set showed that this model can reduce the localization failure rate effectively.3)As for the precise localization problem,we have presented a cascaded regression model,which shares the parameters among all iterations.After concatenating the coarse and precise localization models,we achieved the state-of-the-art result on 300-W test set.Moreover,this paper also points out that the regression model can be interpreted as a gradient prediction model.The experimental results showed that the mean error rate can be further reduced by introducing some techniques in gradient descent method.更多还原