Current Issue Cover

伍凯, 朱恒亮, 郝阳阳, 马利庄(上海交通大学计算机科学与工程系, 上海 200240)

摘 要
目的 人脸配准是当前计算机视觉领域的研究热点之一,其目的是准确定位出人脸图像中具有语义特征的面部关键点,这也是人脸识别、人脸美化等众多与人脸有关的视觉任务的重要步骤。最近,基于级联回归的人脸配准算法在配准精度和速度上都达到了最先进的水准。级联回归是一种迭代更新的算法,初始脸形将通过多个线性组合的弱回归器逐渐逼近真实的人脸形状。但目前的算法大多致力于改进学习方法或提取具有几何不变性的特征来提升弱回归器的能力,而忽略了初始脸形的质量,这极大的降低了它们在复杂场景下的配准精度,如夸张的面部表情和极端的头部姿态等。因此,在现有的级联回归框架上,提出自动估计初始形状的多姿态人脸配准算法。方法 本文算法首先在脸部区域提取基于高斯滤波一阶导数的梯度差值特征,并使用随机回归森林预测人脸形状;然后针对不同的形状使用独立的级联回归器。结果 验证初始形状估计算法的有效性,结果显示,本文的初始化算法能给现有的级联回归算法带来精度上的提升,同时结果也更加稳定;本文算法产生的初始形状都与实际脸型较为相近,只需很少的初始形状即可取得较高的精度;在COFW、HELEN和300W人脸数据库上,将本文提出的多姿态级联回归算法和现有配准算法进行对比实验,本文算法的配准误差相较现有算法分别下降了29.2%、13.3%和9.2%,结果表明,本文算法能有效消除不同脸型之间的干扰,在多姿态场景下得到更加精确的配准结果,并能达到实时的检测速度。结论 基于级联回归模型的多姿态人脸配准算法可以取得优于现有算法的结果,在应对复杂的脸形时也更加鲁棒。所提出的初始形状估计算法可以自动产生高质量的初始形状,用于提升现有的级联回归算法。
Cascade regression based multi-pose face alignment

Wu Kai, Zhu Hengliang, Hao Yangyang, Ma Lizhuang(Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai 200240, China)

Objective Face alignment is one of the most active fields of computer vision. It attempts to localize facial semantic landmarks from a given face image, which is an important step in many face-related vision tasks such as face recognition and face beautification. Cascade regression-based face alignment algorithms have recently achieved state-of-the-art performance in both accuracy and speed. Cascade regression is an iterative method that refines an initial face shape through many linearly combined weak regressors. However, most previous methods focused on boosting the learning method or extracting geometric invariant features while ignoring the initial shape quality. This approach severely lowers their accuracy on complex scenarios, such as exaggerated expressions or extreme head poses. This study proposes a cascade regression-based multi-pose face alignment algorithm initialized with estimated initial shapes. Method The proposed method consists of two parts. First, the first derivative of Gaussian filter-based gradient difference features is extracted to represent the facial appearance, and a random regression forest is learned to predict initial face shapes. Second, these initial shapes are regressed by particular cascade regressors separately. Result The alignment error of this method decreased by 29.2%, 13.3%, and 9.2% in the COFW, HELEN, and 300 W databases, respectively, unlike existing methods. Experiments show that this method can eliminate the disturbance among different shapes for more accurate multi-view face alignment and run in real-time. Conclusion This study proposes an algorithm for multi-pose face alignment based on cascade regression. This algorithm surpasses many state-of-the-art methods in terms of accuracy and is more robust for complex face shapes. The proposed initial shape estimation algorithm can generate initial shapes in suitable quality applied to improve existing cascade regression-based methods.