计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (14): 105-113.DOI: 10.3778/j.issn.1002-8331.2201-0309

• 模式识别与人工智能 • 上一篇    下一篇

基于改进YOLO-v4的室内人脸快速检测方法

巢渊,刘文汇,唐寒冰,马成霞,王雅倩   

  1. 江苏理工学院 机械工程学院,江苏 常州 213000
  • 出版日期:2022-07-15 发布日期:2022-07-15

Fast Indoor Face Detection Method Based on Improved YOLO-v4

CHAO Yuan, LIU Wenhui, TANG Hanbing, MA Chengxia, WANG Yaqian   

  1. School of Mechanical Engineering, Jiangsu University of Technology, Changzhou, Jiangsu 213000, China
  • Online:2022-07-15 Published:2022-07-15

摘要: 针对室内安防工程应用中检测人脸角度不同、光照变化、部分遮挡、模糊等复杂工况,提出一种基于改进YOLO-v4的室内人脸快速检测方法。基于深度可分离残差网络结构改进YOLO-v4主干网络,提升模型检测效率;在构建特征金字塔过程中引入注意力机制,自适应调整通道特征与空间特征权重,提升模型特征提取能力。实验结果表明,该方法对室内人脸图像的检测精度与速度分别为92.53%与35?frame/s,相比原YOLO-v4算法及其他主流人脸检测算法,具有更好的检测精度与效率,因此可应用于移动机器人的室内人脸实时检测。

关键词: 深度学习, 特征融合, 人脸检测, 注意力机制

Abstract: A fast indoor face detection method based on improved YOLO-v4 is proposed in this paper, under the complex working conditions of human faces with different angles, varying illumination, partial occlusion, and blurring in indoor security engineering applications. The YOLO-v4 backbone network is improved based on the deep separable residual network structure to increase the detection efficiency of the model. The attention mechanism is introduced during the process of constructing the feature pyramid, which can adaptively adjust the weights of the channel features and spatial features, to improve the feature extraction capability of the model. The experimental results show that the accuracy and speed of the proposed method are 92.53% and 35 frame/s, respectively, for indoor face images, which has relatively better detection precision and efficiency, compared with the original YOLO-v4 algorithm, and other mainstream face detection algorithms. The proposed method therefore can be applied to indoor face detection of mobile robots in real time.

Key words: deep learning, feature fusion, face detection, attention mechanism