【作者】 彭振云；

【导师】 徐光祐； 艾海舟；

【作者基本信息】 清华大学 ， 计算机应用， 2002， 博士

【摘要】 计算机人脸识别是一个复杂和困难的问题，其原因是：（1）人脸是由复杂的三维曲面构成的可变形体，难以用数学描述；（2）所有人的人脸结构高度相似，而人脸的图像又受年龄和成像条件的影响，这使得同一人在不同条件下的差别可能比不同的人在相同条件下的差别还要大。因此，人脸识别技术实用化所需解决的重要问题是研究能在上述变化条件下可靠工作的人脸识别技术，即鲁棒的人脸识别技术。实现鲁棒的人脸识别涉及人脸检测、特征检测、人脸描述、建模、识别等技术，而其中最关键的是特征检测。基于上述理解，本论文以鲁棒的人脸识别为目标，以人脸特征检测为重点进行了相关的研究，并取得了如下创新性成果：1、提出多线索自适应人脸特征检测方法，将多种人脸线索通过导引、校验、纠错等方式相融合，实现了在姿态、背景和光照变化的情况下人脸特征的可靠检测。与现有典型的特征检测方法相比，正确率和适应性有显著提高（对于姿态变化的情况，正确率提高10%左右），从而使人脸特征检测技术达到实用阶段。2、提出图像分析和运动分析相结合的交叉验证方法，实现了活动图像人脸特征检测中的自动纠错和特征估计机制，从而使视频中人脸特征自动检测的正确率达到98%以上，为基于内容检索和视频编码的应用开辟了道路。3、提出基于自适应阈值分割的人脸特征检测方法和基于主分量分析的特征真实性概念，将人脸特征检测方法建立在无先验假设的基础上，克服了现有方法中过分依赖先验阈值、摄像条件或经验参数的缺陷，极大扩展了人脸特征检测的应用范围。4、将特征检测和颜色分析相融合，提出自适应可变模型的人脸检测方法，解决了复杂背景和光照可变条件下的多人脸检测问题。5、提出基于特征的多姿态自动建模方法，结合多模板匹配策略，有效解决了大容量可变姿态的人脸识别问题。更多还原

【Abstract】 It is difficult to implement the face recognition mechanism using computers for several reasons. First, human face is a deformable object composed of complex 3D curve surfaces, which is hard to be represented in form of mathematics. Secondly, faces of different persons have the similar structure. On the other side, the face images are greatly dependent on ages and photography conditions. This results that the difference between two face images of two different persons taken under the same photography condition is probably less than that between two images of the same person under different conditions. Therefore, in order to implement a practical face recognition system, the recognition method needs to be independent of the above variations. This kind of method is named as Robust Face Recognition technology in this study. A Robust Face Recognition system takes use of many technologies among which feature detection is the most important. According to the above understanding, this study targets on Robust Face Recognition based on feature detection, and makes the following creative contributions:1. An adaptive multi-cue based facial feature detection approach is presented. By combining various facial cues through inducing, verifying and error-correcting, this approach enables to detect facial features robustly under varying poses, background and lightning conditions. It increases the correction ratio significantly (up to 10% for pose variation) compared with typical existing methods and makes it possible to detect facial features for practical use.2. A cross-verification mechanism is worked out for detecting facial features on moving pictures by merging the technologies of both image analysis and motion analysis. It provides a solution to the problem of error-correction and feature estimation and makes it possible to detect facial features on video segment at a correction ration of 98%, thus paves the way for the application of content-based retrieval and video encoding.3. The concept of Feature Fidelity is defined based on Principal Component Analysis (PCA) and an adaptive threshold based feature detector is then proposed, which takes no advantage of ad hoc knowledge including threshold, photography condition, experimental coefficients, etc., thus greatly expands the applications of facial feature detection technology.4. An adaptive face detection method is proposed, which combines feature detection and color analysis and makes it possible to detect multiple faces on the complex background under varied lightning conditions.5. A study on face recognition for varied poses and large face database is made on the basis of automatic multi-view modeling and multi-template matching.更多还原