基于两阶段定位模型的人脸对齐算法研究

发布时间：2018-07-12 14:10

本文选题：人脸对齐 + 人脸关键点定位　；参考：《浙江大学》2017年硕士论文

【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。
[Abstract]:Face alignment is one of the classical problems in computer vision. Its purpose is to automatically calculate the coordinates of the key points in a given face image. Accurate location of key points is of great significance to many visual tasks, such as 3D face reconstruction, facial expression analysis, face pose estimation and so on. With the development of related technology, the current human face alignment method can achieve low positioning error under controlled conditions. However, the input of many human face related applications is obtained under natural conditions. Due to the existence of illumination, background, face pose, image quality and other interference factors, the problem of face alignment is still very challenging. The main contributions of this paper are as follows: (1) through experimental analysis, it is found that reasonable initial values can significantly reduce the localization error rate of cascaded regression models. Based on this discovery, a two-stage face alignment algorithm from coarse to fine is proposed in this paper. Face alignment is divided into two sub-problems: coarse location and fine location, and each problem should be solved by a special method. (2) aiming at the rough location problem, A model based on deep convolution neural network is designed and implemented in this paper. The model takes the whole face as input and directly predicts the position coordinates of the key points of all faces. The results on 300-W test set show that the model can effectively reduce the failure rate of face location at key points. (3) aiming at the problem of precise localization, a cascaded regression model based on parameter sharing is proposed in this paper. Each regression step in the model uses the same parameters. Combined with the rough localization model, the error rate on the 300-W test set is reduced to state-of-the-art. In addition, it is pointed out that a single regression model can be used as a gradient prediction model. By combining the techniques of gradient descent algorithm, the experiments show that the error rate of the key point location can be further reduced.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】