基于图像的人脸特征提取与发型分类

发布时间：2019-03-16 13:35

【摘要】：人脸特征提取是人脸图像分析技术的关键,它被广泛应用于人脸识别、人脸表情分析、三维人脸重建等领域,然而人脸特征点定位不准确的问题依然存在。头发在人体表观中具有重要作用,由于缺少有效的头发分类技术,大大降低了三维人脸变形、3D试衣的真实性。因此,本文针对上述问题进行了相关研究,研究内容主要包括:(1)人脸存在性及合理性的检测。利用肤色检测算法对经过预处理后的图像进行处理,增强图像中的人脸区域,通过协方差矩阵在人脸区域进行坐标系重建,根据其中心点位置构建阈值关系,构建人脸(包含脖颈)区域的最小外接包围盒,根据阈值关系进行人脸存在性及合理性的判定。(2)人脸特征提取。为了获取测试图像中的人脸特征数据,使用主动形状模型(ASM)算法进行处理,针对原算法直接采用灰度值信息构建局部轮廓模型,灰度值对外部自然因素比较敏感,本文采用特征点及其正负方向一定区域像素点的边缘结构方向构建局部轮廓模型,该方法在特征点及相关像素点的法线方向上采集图像的灰度分布特征,能有效利用图像信息。在ASM搜索过程中,利用基于混合投影的眼睛精确定位方法实现眼睛区域的精确定位,根据眼睛位置信息确定模板形状的平移、旋转和尺度变化参数。然后,通过改进的搜索策略进行特征点匹配,进一步提升特征点匹配的准确度。(3)发型分类。提出一种基于颜色空间与曲线变化的能量模型实现对头发区域的分割,根据轮廓检测的图像数据,逐点对头发边缘进行检测分割。利用主成分分析(PCA)和支持向量机(SVM)方法实现对正面图像的发型分类。同时,通过轮廓曲线上各个像素点的曲率定义弯曲度,提出基于弯曲度的局部搜索算法实现对侧面头发长度的分类,然后,根据曲线曲率和经验距离的局部区域搜索方法实现对扎辫子发型的分类。通过实验,验证了以上方法的可行性,实验结果表明,改进的ASM方法取得了较好的特征点定位结果,而本文的发型分类算法也得到了较好的发型分类结果。最后,本文分析了在人脸特征提取和发型分类方面存在的不足并提出了进一步的研究方向。
[Abstract]:Face feature extraction is the key technology of face image analysis. It is widely used in face recognition, facial expression analysis, 3D face reconstruction and so on. However, the problem of inaccurate location of facial feature points still exists. Hair plays an important role in the appearance of human body. Due to the lack of effective hair classification technology, the 3D face deformation is greatly reduced, and the authenticity of 3D fitting is greatly reduced. Therefore, this paper studies the above problems, including: (1) the detection of the existence and rationality of face. The skin color detection algorithm is used to process the pre-processed image, enhance the face region in the image, reconstruct the coordinate system in the face region by covariance matrix, and construct the threshold relation according to the position of its center point. The minimal outer bounding box of the face (including neck) region is constructed to determine the existence and rationality of the face according to the threshold relation. (2) face feature extraction. In order to obtain the face feature data in the test image, the active shape model (ASM) algorithm is used to process it. Aiming at the original algorithm, the gray value information is directly used to construct the local contour model, and the gray value is sensitive to the external natural factors. In this paper, the local contour model is constructed by using the edge structure direction of the feature points and their positive and negative directions, and the gray distribution features of the images are collected in the normal direction of the feature points and related pixels, which can make use of the image information effectively. In the process of ASM search, the precise eye location method based on hybrid projection is used to locate the eye region accurately. According to the eye position information, the parameters of translation, rotation and scale change of the template shape are determined. Then, the accuracy of feature point matching is further improved by the improved search strategy. (3) hairstyle classification. An energy model based on color space and curve change is proposed to segment the hair region. According to the image data of contour detection, the edge of hair is detected and segmented point by point. Principal component analysis (PCA) and support vector machine (SVM) (SVM) are used to classify the hairstyle of positive images. At the same time, the curvature of each pixel on the contour curve is defined by curvature, and a local search algorithm based on curvature is proposed to realize the classification of the length of the side hair. According to the curvature of curves and empirical distance of the local region search method to achieve the classification of braided hair. The feasibility of the above method is verified by experiments. The experimental results show that the improved ASM method has achieved better feature point location results, and the hairstyle classification algorithm in this paper has also obtained better hairstyle classification results. Finally, the shortcomings of face feature extraction and hairstyle classification are analyzed and further research directions are proposed.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】