视频图像中人脸自动检测与统计算法

针对视频中人脸检测由于成像角度、天气状况、遮挡等因素造成检测准确率偏低以及深度学习模型计算复杂度高的问题,文中提出了一种基于椭圆肤色模型与AdaBoost的人脸检测算法。算法通过选取Haar-like特征作为弱分类器,以裁剪过的CAS_PEAL数据集中的人脸图像作为训练集,利用AdaBoost算法将多个弱分类器组合成一个强分类器,最后将若干强分类器以级联的结构组成最终的分类器模型。为解决将非人脸区域检测为人脸的问题,引入椭圆肤色模型,利用椭圆肤色模型对视频帧进行处理使得图像中与肤色相似的区域进入后续的人脸检测过程以降低误检率。实验结果表明,算法能以平均26 ms(单人脸视频)和平均34 ms(多人脸视频)的检测速度进行实时的人脸检测,且达到了87.2%的检测准确率,具有较大的应用推广价值。

Design and Implementation of Automatic Face Detection and Statistics Software in Video Image

Aiming at the problem of low detection accuracy and high computational complexity of deep learning models due to factors such as imaging angle, weather conditions, and occlusion in face detection in video, a face detection algorithm based on ellipse skin color model and AdaBoost was proposed. The algorithm selectd Haar-like features as weak classifiers, and used the face images in the cropped CAS_PEAL data set as the training set, and used the AdaBoost algorithm to combine multiple weak classifiers into a strong classifier. The cascaded structure constituted the final classifier model. In order to solve the problem of detecting non-face area as a human face, an ellipse skin color model was introduced, and the video frame was processed using the ellipse skin color model so that areas with similar skin color in the image enterd the subsequent face detection process to reduce the false detection rate. Experimental results showed that the algorithm could perform real-time face detection at an average detection speed of 26 ms (single face video) and an average of 34 ms (multi-face video), and achieved a detection accuracy of 87.2%, which had a large application promotion value.