基于改进YOLOv5s-face的Face5系列人脸检测算法

Face5 series face detection algorithm based on improved YOLOv5s-face

针对人脸检测中小尺度人脸和遮挡人脸的漏检问题,提出了一种基于改进YOLOv5s-face(you only look once version 5 small-face)的Face5系列人脸检测算法Face5S(face5 small)和Face5M(face5 medium)。使用马赛克(mosaic)和图像混合(mixup)数据增强方法,提升算法在复杂场景下检测人脸的泛化性和稳定性;通过改进C3的网络结构和引入可变形卷积(DCNv2)降低算法的参数量,提高算法提取特征的灵活性;通过引入特征的内容感知重组上采样算子(CARAFE),提高多尺度人脸的检测性能;引入损失函数WIoUV3(wise intersection over union version 3),提升算法的小尺度人脸检测性能。实验结果表明,在WIDER FACE验证集上,相较于YOLOv5s-face算法,Face5S算法的平均mAP@0.5提升了1.03%;相较于先进的人脸检测算法ASFD-D3(automatic and scalable face detector-D3)和TinaFace,Face5M算法的平均mAP@0.5分别提升了1.07%和2.11%,提出的Face5系列算法能够有效提升算法对小尺度和部分遮挡人脸的检测性能,同时具有实时性。

To address the missed detection of small-scale faces and occluded faces in face detection,this paper proposes the Face5 series of face detection algorithms,Face5S(face5 small)and Face5M(face5 medium),based on improved you-only-look-once version 5 small-face(YOLOv5s-face).First,Mosaic and Mixup data enhancement methods are employed to improve the generalization performance and stability of the algorithm to detect faces in complex scenes.Second,the number of parameters of the algorithm is reduced by improving the network structure of C3 and introducing deformable convolution(DCNv2)to improve the flexibility of the algorithm in extracting features.Then,the content-aware reorganization up-sampling operator for features(CARAFE)is introduced to achieve the improvement of the detection performance of the multiple-scale face.Finally,the loss function wise Intersection over union version 3(WIoUV3)is introduced to achieve the improvement of the detection performance of the small-scale face.Our experimental results show on the WIDER FACE verification set,compared with the YOLOv5s-face algorithm,the average mAP@0.5 of the Face5S algorithm rises by 1.03%;compared with the advanced face detection algorithm automatic and scalable face Detector-D3(ASFD-D3)and TinaFace,the average mAP@0.5 of the Face5M algorithm increases by 1.07%and 2.11%respectively.Our proposed Face5 series of algorithms effectively improve the detection performance of the algorithm for small-scale and partially occluded faces with effective real-time performance.