基于改进注意力机制的口罩人脸识别算法

Mask Face Recognition Algorithm Based on Improved Attention Mechanism

口罩人脸识别在新冠肺炎疫情爆发后成为人脸识别领域的难点和热点,传统基于卷积神经网络人脸识别算法由于没有大面积遮挡的去噪能力在口罩人脸识别任务中准确率较低。为了解决以上问题,提出一种基于改进注意力机制的口罩人脸识别算法。在注意力机制引入使用关键点注意力替代全连接操作,形成区块注意力效果,再通过掩码(mask)操作,抑制大面积口罩信息表达并增强有效人脸部分特征,从而提升口罩人脸识别准确率。将改进的注意力机制嵌入ResNet50中,使用上述模型进行人脸特征提取并识别。实验结果表明,在都使用混合数据集(含正常人脸和口罩人脸)训练下,所提算法比ArcFace在LFW和LFW_Mask上准确率分别提高了3.87%和7.91%。

Mask face recognition has become a difficult and hot spot in the field of face recognition after the outbreak of the new crown pneumonia epidemic.Traditional face recognition algorithms have low accuracy in face recognition tasks due to the lack of denoising ability of large area occlusion.In order to solve this problem,a face recognition algorithm for the mask based on the improved attention mechanism was proposed.In the attention mechanism,the key point attention was introduced to replace the full connection operation to form a block attention effect,and then through the mask operation,the large-area mask information expression is suppressed and the partial features of the effective face are enhanced,thereby enhancing the mask face recognition accuracy rate.The improved attention mechanism was embedded in the convolutional neural network ResNet50,and the algorithm was used for facial feature extraction and face recognition.The experimental results show that the accuracy of this algorithm is 3.87%and 7.91%higher than ArcFace on LFW and LFW Mask,respectively,under training using a mixed data set(including normal faces and mask faces).