基于YOLO网络的人脸检测方法

A Face Detection Method Based on YOLO Network

为了提高传统人脸检测方法的准确率,本文使用一种基于Y0L0的人脸检测方法。首先,利用数据增强的方式,从有限的样本中获得大量有效的训练、测试数据;其次,归一化输入图像的尺度,以回归的方式利用单一的网络对整张图片做一次评估,得到目标边界框和类别;最后,根据卷积层中不同大小感受野的特点,通过调整网络的卷积层数量、卷积核大小和池化尺度,最终得到理想的检测模型。将数据增强后获得的正负样本送入构架好的网络模型中进行实验,在评价环节中引入精准率、召回率、F1分数和R0C曲线等指标进行定量分析,在相同的实验环境下与其它几种模型的检测结果进行对比。结果表明,改进的网络在保证检测速率的同时极大地提高了准确率,在各项指标上均有优秀的表现。本文的深度学习人脸检测网络,在保证CNN强大的特征提取能力的同时,继承了Y0L0算法快速高效的特点,面对不同姿态、不同角度、不同光照强度的人脸数据均能取得较好的检测效果,具有较强的实时性、稳定性和鲁棒性。

In order to improve the aeeuraey of traditional faee detection methods, a faee detection model based on YOLO is put forward in this paper. Firstly,a great quantity of effective training and testing data are obtained from limited samples through data enhancement. Then the seale of tire input image is normalized and a single network is used to evaluate the whole picture in order to obtain the border frame and category of tlie object. According to the characteristics of different size of receptive fields in the convolution layer, and ideal detection model is finally obtained by adjusting the number of convolution layers, the size of convolution kernel and the seale of pool. Tlie positive and negative samples with data enhancement are put into a constructed network model for experiment. Indexes including the accuracy rate, recall rate, FI score and ROC curve are introduced into the evalua-tion process for quantitative analysis. Tlie results are compared with those of oilier models in the same experimental condition, and it is shown that the improved model can greatly improve tlie detection rate and accuracy, and it performs well in all indicators. Tlie improved faee detection network model based on deep learning not only ensures tlie powerful feature extraction ability of CNN, but also maintains the fast and efficient characteristics of YOLO algorithm. Besides, it can achieve better detection results with data of different attitudes, angles and light intensity, and excels in real-time performance, stability and robustness.