基于深度学习的自然场景下多人脸实时检测

Multi-Face Real-time Detection in Natural Scene Based on Deep Learning

为了解决人脸检测的实时性与有效性,提出了基于YOLOv3算法的人脸检测模型.该模型以Darknet 53为主干网络,用3种不同尺寸的特征图进行预测,对Bounding box的中心坐标、置信度以及类别损失函数进行设计,最后直接回归被检测人脸的信息.实验中对数据进行了批量归一化处理,加速了Loss收敛.实验采用Wider Face的自然场景下的人脸数据集,将YOLOv3算法模型与不同算法模型比较,结果显示基于YOLOv3算法的人脸检测模型能够保证人脸检测的实时性,同时实现了自然场景下多人脸检测的任务.

In order to solve the real-time and effectiveness of face detection,a face detection model based on YOLOv3 algorithm is proposed by this paper.As the backbone network,Darknet 53 is used in the model which predicts with three different size feature maps and designs the center coordinates,confidence and category loss function of the Bounding box.Finally,it returns the information of the detected face directly.In the experiment,the data is normalized in batches,which accelerates the convergence of Loss.The face dataset in the natural scene of Wider Face is used in this experiment to compare the YOLOv3 algorithm model with different algorithm models.The results show that the face detection model based on YOLOv3 algorithm can guarantee the real-time performance of face detection and realize the multi-person face detection in natural scene.