| 摘 要: 提出一种教室场景下人脸检测与识别的算法，基于RetinaFace人脸检测框架进行改进，在主干网络中引入可变形卷积以适应人脸遮挡以及人脸变形，调整预设Anchor并在上下文敏感模块中引入残差结构以适应教室场景下尺度变化的特点。在公开数据集WIDER FACE上训练基础权重，然后在教室场景下自标注的数据集中进行迁移学习以适应教室场景，最后通过ArcFace人脸识别网络进行人脸识别。本算法在公开数据集WIDER FACE上batch size设置为16时，Easy、Medium和Hard的人脸检测精度分别为96.34%、95.12%和89.64%；在自标注的数据集上batch size设置为4时，人脸检测精度为94.72%，人脸识别精度为92.11%。实验结果表明，该算法可以有效提高教室场景下人脸检测与识别的效果。 |
| 关键词: 人脸检测；人脸识别；RetinaFace；ArcFace |
|
中图分类号: TP399
文献标识码: A
|
| 基金项目: 教育部科技项目(NO.2020HYA02007)资助. |
|
| Face Detection and Recognition in Classroom Scenarios |
|
YANG Kaiwen1, YANG Jiale2
|
( 1.School of Information, Shenyang University of Technology, Shenyang 110870, China ;
2.School of Artificial Intelligence, Shenyang University of Technology, Shenyang 110870, China )
527410159@qq.com; 2082873733@qq.com
|
| Abstract: This paper proposes an algorithm for face detection and recognition in classroom scenes, which is improved based on the RetinaFace face detection framework. Deformable convolution is introduced into the core network to adapt to face occlusion and face deformation. The preset Anchor is adjusted and the residual structure is introduced into the contextsensitive module to adapt to the characteristics of scale changes in classroom scenes. The basic weights are trained on the public dataset WIDER FACE, and then the self-labeled dataset in the classroom scene is used for transfer learning to adapt to the classroom scene. Finally, face recognition is carried out through ArcFace face recognition network. When the batch size of this algorithm is set to 16 on the public dataset WIDER FACE, the face recognition precision rates of Easy, Medium and Hard are 96.34%, 95.12% and 89.64% respectively. When the batch size on the self-labeled dataset is 4, the face detection precision is 94.72%, and the face recognition precision is 92.11%. Experimental results show that the algorithm can effectively improve the effect of face detection and recognition in classroom scenes. |
| Keywords: face detection; face recognition; RetinaFace; ArcFace |
