牛脸特征点检测的研究与实现

发布时间：2018-07-24 11:19

【摘要】：牛的面部轮廓是牛脸的重要特征,牛脸特征点的检测及其形状分析可以用于牛身份鉴别、咀嚼分析及健康状况评估等领域。针对真实生产环境下牛场视频监控图像中存在的拍摄角度差异大、光照不均匀、牛脸局部有遮挡等问题。本文研究了监督式梯度下降算法(SDM),局部二值算法(LBF)和主动外观模型算法(FAAM)三种算法提取牛脸轮廓信息,为进一步分析牛的面部表情和健康状况提供了理论基础。主要研究内容和结论如下:(1)牛脸检测器训练。基于AdaBoost级联分类器原理,结合牛脸的特征,训练牛脸检测器。由于牛脸较长,截取牛脸正例图像大小为15×25像素。负例样本是与正例样本尺寸相同的牛脸背景图像。最终训练出的牛脸检测器在600幅单张牛脸图像中,检测率达到了93%。(2)牛脸特征点标定研究。根据特征点选取的规则,选择29个点来标记牛脸轮廓,并实现了特征点标定软件。手动标记600幅牛面部图像,每幅图像的特征数据保存成文本格式。对所有的训练集进行对齐处理,对齐后的数据集为后续模型的建立提供数据。(3)三种特征点检测算法在牛脸特征点检测中的应用。牛脸特征点检测过程中由于牛脸较长,对初始化算法进行了改进,采用分离式模型对牛脸面部特征点进行初始化。通过实验分析和比较了三种算法的时间效率和准确性,并采用三种误差分析方法比较了算法的性能。实验结果表明,LBF算法,SDM算法以及FAAM算法每幅图像的平均检测时间分别为0.39秒,1.48秒,71.34秒。LBF算法,SDM算法以及FAAM算法尺寸归一化点对点的平均误差分别为0.0359像素,0.0275像素,0.0269像素,左右眼角归一化后点对点的欧式距离度量平均误差分别为0.0245像素,0.0188像素,0.0184像素,均方根误差分别为0.0323像素,0.0247像素,0.0242像素。实验结果验证了牛脸特征点检测的可行性和实用性,其中FAAM算法获得的模型精度最高,而LBF算法的计算效率最高。因此在牛脸特征点检测过程中,根据实际需求要在准确率和时间效率上做出取舍。
[Abstract]:The facial contour of cattle is an important feature of cattle face. The detection and shape analysis of bull face feature points can be used in identification, chewing analysis and health assessment of cattle. Aiming at the problems of large angle difference, uneven illumination and partial occlusion of cattle face in the real production environment, the video surveillance image of cattle farm is different. In this paper, the supervised gradient descent algorithm (SDM),) local binary algorithm (LBF) and the active appearance model algorithm (FAAM) are studied to extract bovine face contour information, which provides a theoretical basis for further analysis of cattle facial expression and health status. The main contents and conclusions are as follows: (1) Bovine face detector training. Based on the principle of AdaBoost cascade classifier and the features of bovine face, the bovine face detector is trained. Because the cattle face is longer, the sample image size is 15 脳 25 pixels. The negative sample is a cattle face background image with the same size as the positive sample. Finally, in 600 single bovine face images, the trained bovine face detector has a detection rate of 93%. (2) study on the calibration of bull face feature points. According to the rule of feature point selection, 29 points are selected to mark the bull face contour, and the feature point calibration software is implemented. Manually mark 600 cattle facial images, each image's feature data saved as text format. All training sets are aligned and the aligned data sets provide data for the subsequent modeling. (3) the application of three feature point detection algorithms in cattle face feature point detection. In the process of cattle face feature point detection, the initialization algorithm is improved because of the long cow face, and the split model is used to initialize the cattle face feature point. The time efficiency and accuracy of the three algorithms are analyzed and compared experimentally, and the performance of the three algorithms is compared by using the three error analysis methods. The experimental results show that the average detection time of LBF algorithm and FAAM algorithm is 0.39 seconds, 1.48 seconds, 71.34 seconds, respectively, and the average error of normalized point-to-point of FAAM algorithm is 0.0359 pixels / 0.0275 pixels / 0.0269 pixels, respectively. The average error of Euclidean distance measurement after normalization of left and right eye corners is 0.0245 pixel / 0.0188 pixel / 0.0184 pixel, and the root mean square error is 0.0323 pixel / 0.0247 pixel / 0.0242 pixel respectively. The experimental results verify the feasibility and practicability of the bull face feature point detection. The FAAM algorithm has the highest model accuracy and the LBF algorithm has the highest computational efficiency. Therefore, in the process of bull face feature point detection, we should make a choice between accuracy and time efficiency according to actual demand.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】