基于M-Xception网络的戴口罩人脸表情识别

Facial Expression Recognition with Mask Based on M-Xception Net

目前,国内外对于有遮挡的人脸表情识别研究较少,其中戴口罩的人脸表情识别(faceemotionrecognition withmask,FERM)应用场景复杂、数据集缺乏、识别准确率低,因此提出一种改进的Xception网络模型M-Xception Net(Modified Xception Net),并建立一个FERM数据集.M-Xception网络模型具有轻量级特性,参数量较少,对细微表情信息敏感,适用于场景复杂、分辨率较低表情数据集,并利用Dlib的Face-Mask技术,对FER2013表情数据进行戴口罩的遮挡处理,生成FERM数据集.将M-Xception网络模型在FERM数据集上进行测试,结果表明,戴口罩人脸表情识别准确率达到88.95%,高于直接利用Xception网络进行戴口罩表情识别的准确率84.94%,并且缩短了训练时间.

At present,there is little research on masked facial expression recognition with occlusion at home and abroad.For the masked facial expression recognition(face emotion recognition with mask,FERM)with the complex application scenarios,lack of data set,and low recognition accuracy rate,thus this article proposes an improved M-Xception Net(Modified Xception Net),and establishes a data set FERM.It has lightweight network characteristics with fewer parame-ters,more sensitive to subtle expression information.Therefore,it is more suitable for complex scenes and emoticon datasets of lower resolution.The Face-Mask technology of Dlib is applied to mask FER2013 expression data and generates FERMexpression data set.The experiments on FERM data set show that the M-Xception network model has better recognition per-formance with the 88.95%accuracy,which is higher than the 84.94%accuracy and shorter training time with using Xception net directly.