【作者】 陈莉；

【导师】 周激流；

【作者基本信息】 四川大学 ， 通信与信息系统， 2002， 硕士

【摘要】 肤色是一种简单有效的特征，在人脸检测中得到广泛的关注和应用，但传统的基于统计的肤色检测方法不能克服光照、摄像机、肤色差异等因素的影响。 本文的工作：通过分析遗传算法的交叉和变异算子，引入具有交叉位置控制的最大Hamming距交叉算子以及具有变异位置控制的变异算子，性能测试结果比SGA和AGA都有明显的提高；选择适当的网络结构，使其肤色检测结果对门限选取的敏感程度大大降低，具有良好的稳定性；采用改进的遗传算法和BP算法相结合的方法进行网络的权值和阈值的优化，网络的性能较之单独采用BP算法优化得到明显改善；在肤色检测中，引入数据的预处理环节，大幅提高了网络的肤色检测速度；通过肤色特征和几何特征进行眼睛块的分割，对候选眼睛对编号，将编号作为遗传算法的解空间并实现了人脸的定位。新意之处：利用肤色是一种感知现象的特性以及神经网络的强大感知、学习功能，采用进化神经网络对肤色进行感知和分类；以眼睛的几何和周边的肤色特征分割眼睛块，克服了在完整的肤色区域内检测人脸方法的局限性；对候选眼睛对的编号而不是人脸在图像中的位置、人脸的尺度和方向进行编码，使遗传算法用于人脸定位时的问题空间得以大大简化，采用一个新颖的眼睛及脸颊部位的灰度投影及肤色模板验证，可以实现多尺度、任意方向的正面人脸检测，实验结果令人满意。 第一部分，遗传算法原理、算法改进及其性能测试；传统肤色检测方法的研究。第二部分，基于进化神经网络的肤色检测算法。第三部分，复杂背景彩色图像中用遗传算法实现了人脸定位。更多还原

【Abstract】 Human skin color has been proven to be an effective feature and widely used in face detection in resent years. However, color is not a physical phenomenon. It is a perceptual phenomenon that is related to the spectral characteristics of electromagnetic radiation in the visible wavelengths striking the retina. How to improve the performance of the skin detector is a challenging problem.Neural Networks are parameterized non-linear models used for empirical regression and classification modeling. Their flexibility makes them able to discover more general relationships in data than traditional statistical models.Firstly an improved GAs is proposed and shows better performance in comparison with SGA and AGA. Secondly the traditional statistical skin models are studied and a novel approach to design a neural network based skin detector is put forward, which will be later used to retrieve skin-like pixels in color face images. Further more, an evolutionary search procedure (GAs) also has been introduced into the neural network training process. The experiment shows that the evolutionary neural network based skin detector performs much better than traditional skin color models.At last, we applied both the improved GAs and the skin detector to the face detection, and got impressive results. The eyes can be considered as a salient andrelatively stable feature of faces, so firstly eyes-analogue regions in cluttered images are segmented using local adaptive threshold edge detector. Then the small eye-analogue regions are grouped together and labeled using a traditional labeling process according to their geometrical and color features. Instead of finding potential eye-pairs from eye-analogue regions one by one, all possible pairs of eyes are encode as the solution, and the potential face are searched by GAs. In this way, the possible solution space is reduced dramatically. The evaluate function is defined by the combination of horizontal gray projection and skin color in the regions of eyes and cheek.更多还原