Computer Engineering and Applications ›› 2019, Vol. 55 ›› Issue (4): 173-178.DOI: 10.3778/j.issn.1002-8331.1710-0280

Previous Articles     Next Articles

Research on Deep Convolution Neural Network with Small Filter Used in Facial Landmark Detection

ZHENG Yinhuan1, WANG Beizhan2, WANG Jiajun2, CHEN Lingyu2, HONG Qingqi2   

  1. 1.Department of Computer Science and Engineering, Xiamen Institute of Technology, Xiamen, Fujian 361021, China
    2.Software School, Xiamen University, Xiamen, Fujian 361001, China
  • Online:2019-02-15 Published:2019-02-19

深度卷积神经网络应用于人脸特征点检测研究

郑银环1,王备战2,王嘉珺2,陈凌宇2,洪清启2   

  1. 1.厦门工学院 计算机科学与工程系,福建 厦门 361021
    2.厦门大学 软件学院,福建 厦门 361001

Abstract: To solve the problem in a complex environment, such as different positions, light conditions, and barrier factors which lead to a big drop in precision by using traditional facial feature point detection algorithms, the research on the basis of theoretical knowledge is made, and the deep convolution neural network based on small filter is put forward. The algorithm introduces the small filter thought and depth-first network into deep convolution neural network model, redesigns the training in view of the facial feature points detection, and improves the effectiveness and applicability of the algorithm. By applying the algorithm in ALFW and AFW face datasets to predict five points of facial feature, and compared with several other classical algorithms analysis results, it shows that the deep convolution neural network based on the small filter in the prediction of facial feature points at five issues has better accuracy and robustness.

Key words: small filter, deep convolution neural network, feature point detection, depth of network

摘要: 为解决在复杂环境下,如姿势不同、光照条件以及遮挡等因素导致传统人脸特征点检测算法的精度大幅度下降的问题,在特征点检测理论知识以及研究现状的基础上,针对传统卷积神经网络模型在处理人脸特征点检测问题时的不足之处,提出基于小滤波器的深卷积神经网络。算法引入小滤波器思想和以拓展“网络深度”优先的深层卷积神经网络模型,针对人脸特征点检测重新设计训练,提高了算法的有效性与适用性。通过将算法应用于ALFW和AFW人脸数据集上预测5点人脸特征点问题,并与其他多个经典算法进行对比分析,结果表明:基于小滤波器的深卷积神经网络在预测人脸5点特征点问题上有更好的准确性和鲁棒性。

关键词: 小滤波器, 深卷积神经网络, 特征点检测, 网络深度