基于改进的LBP特征和随机森林相结合的人脸关键点检测方法研究

Research on Face Key Point Positioning Method Based on Random Forests and the Improved LBP Features

提出了一种人脸关键点检测方法,该方法用了少量的正面图像,不用归一化人脸图像,而传统的人脸关键点检测方法需要对图像进行严格预处理。随机森林是一种分类器融合算法,可以很好地解决多类分类问题,虽然LBP特征简单,但其可以包含大量的纹理信息。利用改进的LBP特征与随机森林相结合,构成一种对人脸关键点检测的方法。通过高斯平滑图像的LBP特征的提取,对每个点生成特征,计算出有用的特征作为正例,并且与反例集合变为训练集。通过随机森林分类器进行分类,误差率较低,仅在10%左右。

The paper presents the positioning method of the critical points of a face. The method using a small amount ot the front image, without the normalized face image, but the traditional face critical point positioning the image needs to be strictly pretreatment. Random forest is a classifier fusion algorithm, can be a good solution to the multi-class classification problem, LBP feature is a simple feature, but it can contain a very large number of texture information, this article by im- proved LBP features random forests combined constitute a method to locate key points on the face. First make images with gaussian smoothing algorithm; then use the LBP features extracted for each point to generate the feature, calculate the useful features as positive examples, and the training set with counterexample collection becomes; finally use the random forest classifier to classify good error rate-about 10 %.