基于图推理的人脸遮挡检测方法研究  

Study on Face Occlusion Detection Based on Graph Reasoning

作　　者：司春晖 张丽红[1] SI Chunhui;ZHANG Lihong(College of Physical and Electronic Engineering, Shanxi University, Taiyuan 030006, China)

机构地区：[1]山西大学物理电子工程学院,山西太原030006

出　　处：《测试技术学报》2021年第6期515-520,共6页Journal of Test and Measurement Technology

基　　金：山西省研究生教育改革课题资助项目(2019JG031);山西大学物理电子工程学院课程思政建设资助项目(WDKCSZ202009);山西省研究生创新资助项目(2020SY024)。

摘　　要：卷积神经网络的架构局限性及其计算复杂性影响了遮挡人脸的识别精度,其主要原因为不能判别人脸中具体的遮挡位置,导致在识别存在未知遮挡的人脸时,识别精度会急速下降.本文提出一种基于图卷积推理网络的人脸遮挡位置检测网络,利用残差网络和空间金字塔池化提取人脸的低层、高层及边缘特征,通过图投影将相似的像素特征投影到图节点,计算节点之间的投影数据关系以推理分析可能遮挡的区域,并为该区域分配像素进行检测,最终检测出人脸遮挡区域.采用Helen等数据集进行模型训练和测试,经过实验验证,该方法检测精度和分割精度均优于目前其他神经网络人脸遮挡检测的方法.The architecture limitation and computational complexity of convolutional neural network affect the recognition accuracy of occluded face.The main reason is that it can't identify the specific occluded position in the face,which leads to the rapid decline of recognition accuracy when recognizing the face with unknown occlusion.In this paper,we propose a face occlusion detection network based on graph convolution inference network.Residual network and spatial pyramid pooling are employed to extract low-level,high-level and edge features of human face.Similar features of pixels are projected to graph nodes by graph projection,and the projection data relationship among nodes is calculated to infer and analyze the possible occlusion areas.After that this area is assigned pixels for detection.Finally,the occluded area is detected.Helen and other data sets are used for model training and testing.Experimental results show that the detection accuracy and segmentation accuracy of this method are better than other face occlusion detection methods based on neural network.

关 键 词：遮挡检测 人脸分割 图推理 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]