Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (11): 242-249.DOI: 10.3778/j.issn.1002-8331.2011-0162

• Graphics and Image Processing • Previous Articles     Next Articles

Light-Weight Anchor Free Face Detection Based on Multi-Feature Fusion

HUANG Siwei, LI Zhidan, CHENG Jixiang, LIU Andong   

  1. School of Electrical Engineering and Information, Southwest Petroleum University, Chengdu 610500, China
  • Online:2022-06-01 Published:2022-06-01



  1. 西南石油大学 电气信息学院,成都 610500

Abstract: In view of the slow detection speed caused by the huge amount of computation and parameters in deep convolutional neural networks and the low accuracy by complex scenes in nowadays face detection, this paper presents an anchor-free face detection algorithm combing light-weight backbone and multi-feature fusion. The method firstly constructs a light-weight convolutional neural network as the backbone for feature extraction. Then three different modules are used to process the feature fusion:receptive field enhancement module to strengthen image information extraction, weighted feature fusion module to improve detection accuracy, channel shuffle fusion module to simplify the detection head. Finally, the center point anchor-free detection method is used to calculate the fused features and predict the position of faces. The experimental results show that the final model parameter is only 5.1?MB. The detection accuracy of easy, medium and hard in WIDERFACE verification set are improved by 1.4, 2.2 and 4.8 percentage points respectively compared with the benchmark method, which shows that the method has high accuracy in complex scenes face detection with a light-weight model and verifies the efficiency of the proposed method.

Key words: face detection, feature fusion, anchor free detection, light-weight neural network

摘要: 针对大多数基于深度卷积网络的人脸检测方法存在因模型参数和计算量大造成的检测速度慢,以及复杂场景下人脸检测准确率低的问题,提出一种基于多特征融合的轻量化无锚人脸检测方法。构造轻量化卷积神经网络作为特征提取的骨干网络,以加速网络计算过程;引入三种模块处理并融合特征层,包括:感受野增强模块强化图片信息提取、权重特征融合模块提升检测准确性以及通道混洗融合模块简化计算过程;使用中心点定位的无锚检测方法对融合后的特征进行预测。实验结果显示,该方法模型参数量仅为5.1?MB,对比该基准方法,在WIDERFACE验证集中的简单、中等和困难难度的检测准确率分别提升1.4、2.2和4.8个百分点,表明该方法在保证模型轻量化的同时对复杂场景人脸有着较高的检测精度,验证了所提方法的有效性。

关键词: 人脸检测, 特征融合, 无锚检测, 轻量化神经网络