基于改进Yolov5模型的实时人脸检测算法

针对疫情背景下，在一些人流密集场所进行体温筛查或身份识别，当待检测对象快速通过时，人脸检测实时性不高的问题，提出了一种改进Yolov5模型的实时人脸检测算法。该算法首先对骨干网络层进行轻量化改进并引入注意力机制减少冗余信息；其次修改了检测层网络结构，增加了对小目标人脸及倾斜人脸检测的适应性；随之使用Focal EIOU损失函数代替Yolov5原始损失函数中的GIOU损失函数来计算定位损失，有效解决了预测框在目标框内部或预测框与目标框大小一致时无法精确定位的问题。实验结果表明：提出的实时人脸检测算法检测精度达到97.2%,检测速度达到66.7 f/s,相较于原始Yolov5算法，检测精度提升了19.7%,检测速度提升了24 f/s,满足实时人脸检测要求，同时对于黑暗环境及不同表情姿态人脸检测也有较好的适应性。

Real time face detection algorithm based on improved Yolov5 model

Aiming at the problem of poor real time face detection when the object to be detected passes quickly in some crowded places for temperature screening or identification in the context of an epidemic,a real time face detection algorithm based on the improved Yolov5 network framework is proposed in this paper.Firstly,the light weighting of the backbone network layer is improved and an attention mechanism to reduce redundant information is introduced.Secondly,the network structure of the detection layer is modified to increase the adaptability for small target face and tilted face detection.Finally,by using the Focal EIOU loss function replaces the GIOU loss function in the original loss function of Yolov5 to calculate the localization loss,which effectively solves the problem that the prediction frame cannot be accurately,positioned when the prediction frame is inside the target frame or when the prediction frame is the same size as the target frame.The experimental results show that the proposed real time face detection algorithm achieves 97.2% detection accuracy and 66.7 f/s detection speed,which is 19.7% improvement in detection accuracy and 24 f/s improvement in detection speed compared with the original Yolov5 algorithm,meeting the real time face detection requirements,and also has better adaptability for dark environment and different expression pose face detection.