计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (10): 1727-1734.DOI: 10.3778/j.issn.1673-9418.1911037

• 人工智能 • 上一篇    下一篇

针对二分支神经网络匹配的人脸检测算法研究

王茹,贺兴时,杨新社   

  1. 1. 西安工程大学 理学院,西安 710600
    2. 密德萨斯大学 科学与技术学院,英国 剑桥 CB2 1TN
  • 出版日期:2020-10-01 发布日期:2020-10-12

Research on Face Detection Algorithm for Dual-Shot Neural Network Matching

WANG Ru, HE Xingshi, YANG Xinshe   

  1. 1. College of Science, Xi??an Polytechnic University, Xi??an 710600, China
    2. College of Science and Technology, Middlesex University, Cambridge CB2 1TN, UK
  • Online:2020-10-01 Published:2020-10-12

摘要:

在尺度、遮挡等因素的影响下,人脸检测算法速度和精度不匹配进而表现出检测性能差等系列缺点,采用融入模板匹配思想的神经网络检测器算法解决这一不足,提出了二分支神经网络匹配的人脸检测算法。以此得到检测速度与精度相匹配的改进方法。首先根据五官位置关系构建人脸模板框架并设置一定的阈值;其次依据构建的模板和锚策略对不同尺度人脸的鲁棒性,在DSFD人脸检测器的特征增强模块层中实现对样本数据中人脸区域的提取及标记;最后依据提取、标记的人脸区域与构建的模板进行相关性匹配。改进算法在Face Detection Data Set and Benchmark、FDDB、WIDER FACE人脸数据集的实验结果表明,在提升检测速度的同时保证算法的精度,相比几类相近算法,该算法的优势更加明显。

关键词: 人脸检测, 模板匹配, 人脸检测器, 特征提取

Abstract:

Under the influence of scale, occlusion and other factors, the speed and precision of the face detection algorithm are not matched at the same time, and the detection performance is poor. The neural network detector algorithm incorporating the idea of template matching is adopted to solve this deficiency. A face detection algorithm for dual-shot neural network matching is proposed. Thus, an improved method for matching the detection speed and precision is obtained. Firstly, constructing a face template framework and setting a certain threshold according to the five-feature position relation. Secondly, according to the robustness of the constructed template and anchor strategy to different scales of face, extracting and labeling the face area in the sample data in the feature enhancement module layer of the DSFD (dual shot face detector). Finally, matching the correction between the extracted and marked face area and the constructed template. The experimental results on face data sets, such as Face Detection Data Set and Benchmark, FDDB and WIDER FACE indicate that the improved algorithm ensures the accuracy while improving the detection speed, and its advantage is more obvious than that of several similar algorithms.

Key words: face detection, template matching, face detector, feature extraction