基于卷积特征的人脸特征点定位研究

发布时间：2018-03-03 02:31

本文选题：人脸特征点定位　切入点：非可控场景　出处：《北京交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：人脸特征点定位是自动定位事先定义好的面部关键点,以获取人脸形状。人脸特征点定位是人脸图像分析的关键步骤,在人脸识别、仿真、跟踪、表情分析、人脸三维动画建模等任务中有着广泛的应用,受到研究者们广泛的关注。目前,在可控环境下,人脸特征点定位和人脸分析相关任务已经达到了比较满意的结果。但在非控场景下,由于受到光照、姿态、表情、遮挡等非可控因素的影响,人脸外观和形状变化呈现高度非线性。因此,已有模型在人脸特征点定位和人脸其他相关任务上的性能依然很差,无法满足实际应用的要求。本论文主要针对非可控场景下的人脸特征点定位问题进行研究,并针对该问题提出了相应的解决办法。主要创新点如下:(1)在非可控场景下,人脸外观变化呈现高度非线性,而现有模型的视觉特征描述方法难以对其进行准确表达。为了解决该问题,本文提出一种基于深度卷积特征和极限学习机的人脸特征点定位方法。该方法包括:首先,设计并训练深度卷积神经网络,提取全局卷积特征,包含空间像素和上下文语义信息;然后,引入鲁棒的极限学习机,代替卷积神经网络中内置的回归器,实现人脸外观特征到人脸形状的非线性映射;最后,融合多尺度的预测结果确定特征点位置。实验结果表明:包含空间像素和上下文语义信息的卷积特征凸显了复杂人脸外观的一般模式,更有助于定位任务;极限学习机回归函数对人脸外观特征到人脸形状之间的映射能力较强;融合多尺度预测可以进一步提高定位的精度。(2)在非可控场景下,人脸外观和人脸形状存在较大的差异,而现有的级联形状回归模型对初始形状敏感,局部特征忽略形状约束信息。为了解决该问题,本文提出一种改进的级联回归模型实现人脸特征点定位。该模型包括:首先,在级联结构的第一级,通过学习算法直接输出所有特征点位置作为初始形状,代替人工赋值的初始形状,在级联结构的后几级,设计并训练多目标的浅层卷积神经网络,提取局部卷积特征,包含局部相关性信息;然后,改进极限学习机回归函数的优化方法,提高算法的泛化性;最后,通过级联框架实现由粗到精的人脸特征点定位。实验结果表明:通过学习算法初始人脸形状鲁棒性较高;增加局部特征之间的相关性可以添加形状约束信息,提高定位的精度。另外,本文提出的级联形状回归模型每级基于不同类型的特征,拓宽了现有的级联形状回归方法。
[Abstract]:Face feature point localization is the key point of the face that is defined in advance to obtain the shape of the face. Face feature point location is the key step of face image analysis, in face recognition, simulation, tracking, expression analysis, face recognition, simulation, tracking, facial expression analysis, Human face 3D animation modeling and other tasks have been widely used by researchers. At present, in a controllable environment, The tasks related to face feature location and face analysis have achieved satisfactory results. However, in the non-controlled scene, due to the influence of uncontrollable factors such as illumination, pose, facial expression, occlusion, etc. Face appearance and shape change are highly nonlinear. Therefore, the performance of existing models in facial feature location and other related tasks is still very poor. This paper mainly focuses on the problem of face feature point localization in uncontrollable scene, and puts forward the corresponding solution. The main innovation is as follows: 1) in the uncontrollable scene, The appearance change of human face is highly nonlinear, but it is difficult to express it accurately by the visual feature description methods of the existing models. In order to solve the problem, In this paper, a face feature point localization method based on deep convolution feature and extreme learning machine is proposed. The method includes: firstly, the deep convolution neural network is designed and trained to extract global convolution feature. It contains spatial pixels and contextual semantic information. Then, a robust extreme learning machine is introduced to replace the built-in regression in convolution neural network to realize the nonlinear mapping from facial appearance to face shape. The experimental results show that the convolution feature, which contains spatial pixels and context semantic information, highlights the general pattern of complex face appearance, and is more helpful to localization task. The regression function of extreme learning machine has a strong ability to map the appearance features to the shape of the face, and the fusion of multi-scale prediction can further improve the accuracy of location. (2) in the uncontrollable scene, there are great differences between the appearance of the face and the shape of the face. However, the existing cascade shape regression model is sensitive to initial shape and local features ignore shape constraint information. In order to solve this problem, an improved cascade regression model is proposed to locate facial feature points. In the first stage of cascade structure, all feature points are directly output as initial shape by learning algorithm, instead of the initial shape of artificial assignment. In the later stages of cascade structure, a multi-objective shallow convolution neural network is designed and trained. Extract local convolution features, including local correlation information; then, improve the optimization method of LLM-regression function, improve the generalization of the algorithm; finally, Face feature point localization from coarse to fine is realized by cascading frame. Experimental results show that the initial face shape is robust by learning algorithm, and shape constraint information can be added by increasing the correlation between local features. In addition, the cascade shape regression model proposed in this paper is based on different types of features, which broadens the existing cascade shape regression methods.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】