【作者】 徐战武；

【导师】 朱淼良；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2006， 博士

【摘要】 皮肤检测是人脸检测与识别、表情识别、手势识别、人体检测等计算机视觉任务的重要组成部分，更是图象与视频索引、色情图象检测的关键步骤，广泛应用于人机交互接口、访问控制、视频监控以及互联网敏感内容过滤等领域。基于颜色的皮肤检测具有简单、快速、直观，不受物体形状变化及视点改变等影响的优势，受到研究者的普遍重视，具有重要的理论研究意义和应用价值。本文主要研究静态图象中的肤色检测技术。 颜色空间的选择和肤色建模方法是肤色检测的关键问题。肤色检测首先面临一个合适颜色空间的选择问题，大多数研究者直观地选择了“最优”颜色空间而没有给出严格的证明，而一些作者则质疑空间选择对皮肤检测结果的明显影响。在系统回顾和综合分析了肤色检测中所采用的各种颜色空间与肤色模型，并按照各自特性进行分类后，定义了基于空间分布本身的肤色内聚性和肤色-非肤色可分离性两类指标，基于一个大肤色样本库，比较了包含所有常用颜色空间以及部分作者提出的最优空间的17个颜色空间的相应指标，并在这些空间中训练了SPM、GMM、SOM和SVM这四种最典型的统计肤色模型并测试了其综合分类性能。同时考察了量化等级、模型精度和不同决策方式对性能的影响，构成了一个完善的性能评价体系，获得了肤色空间和模型选择的综合结论。 在训练的收敛阶段，SOM参考矢量的改变量很小，基于这一事实，我们利用统计直方图构造获胜点线性表，并根据参考矢量的运动方向来优化获胜点的局部搜索，作为全局获胜点的近似。肤色样本实际分布数据和人工合成数据的实验结果表明该局部搜索策略能够有效提高搜索效率。 照明条件、照相机特性以及噪声都对肤色有很大的影响，个体肤色也各不相同的，肤色模型必须能够适应这些环境变化。针对肤色检测，详细介绍了各种光照补偿方法和自适应手段的优缺点，并比较了用于视频序列的动态高斯模型和动态直方图。 人像通常位于成像系统的焦平面上，人体的周围环境将因散焦而模糊，尽管焦平面中的皮肤区域很难保证有完整的比较强的边界，但是总能找到一条或几条具有一定长度的强边界段。我们提出了基于强边界段的焦平面肤色检测框架，包含强边界段或与之邻接的类肤色区域识别为皮肤区域，从而避免了皮肤区域边界的不完整性，并在保证肤色区域内聚性的前提下，从散焦肤色区域中进一步提取出最接近的子区域作为肤色区域。 皮肤区域具有良好的统计内聚性和空间平滑性，结合纹理等邻域特征能够提高皮肤检测性能。最后我们提出了一个多种底层特征相融合，并与人脸、手部等高层目标结合的皮肤检测开放式框架，形成一个多种信息相互印证，互为目的和手段的启发式综合方法。结合人脸检测的实验结果验证了该框架的有效性。更多还原

【Abstract】 Skin detection plays a key role in many computer vision tasks like face detection and recognition, expression recognition, gestures recognition, human detection, content-based image and video indexing, etc, especially in adult image detection. It has been widely used in the area such as human machine interface, access control, surveillance and objectionable Web images filtering system and so on. Skin color has proven to be a powerful cue for skin detection in images because of its advantages: low computational cost and robustness against viewpoint changing and geometrical transformations. This dissertation focuses on detection of skin color in static images.The choice of the color space and the way of modeling the skin color distribution are key problems for skin color detection. We review the commonly used color spaces and skin color models comprehensively, and classify them according their properties. Using a large data set of 1894 images, we examine whether the color space transformation can increase the compactness of skin class and the discriminability between skin and non-skin classes in seventeen color spaces. We also evaluate the classification performance of SPM, GMM, SOM and SVM in these color spaces. The effect of histogram size, dropping illumination, precision of model and decision strategy is analyzed respectively. This new comprehensive color space and color model testing methodology would allow for making the best choices for skin detection in general.The SOM changes only gradually during its final fine-tuning phase, the new winner of same training input may be found at or in the vicinity of the old one. We propose a novel local search strategy based on movement of surroundings weight vector to accelerate winner search. Experiments with artificial and real world data showed that the local search algorithm is noticeably better in performance than the conventional one.Due to variations of lighting conditions, camera hardware settings, and the range of skin coloration among human beings, a predefined skin color model cannot accurately capture the wide distribution of skin colors in individual images. The most common used illuminant compensation and adaptation approach, dynamic model including dynamic histogram and dynamic Gaussian model are reviewed in detail in chapter 4.The human is usually well focused when captured by the lens system, whereas background objects are typically blurred to out-of-focus. It is difficult to find a closed boundary for focused skin regions. We propose a skin color region detection solution is based on salient boundary segment. A skin-like region containing or adjoining salient boundary segment is regard as skin region. Those defocused skin-like regions更多还原