生成式方法在戴面罩热红外人脸识别中的研究

Research of Generative Method in Thermal InfraredFace Recognition with Face Mask

随着人工智能技术的发展,可见光人脸识别大放光彩。但是可见光无法解决面部有遮挡的问题,因此限制了它在国家安全方面的应用范围。因为可见光无法透过遮挡物从而被相机接收,所以丢失了大量信息。现实中如果佩戴面罩,鼻子、嘴巴、脸颊等大部分的特征将会丢失,则会导致可见光人脸识别的识别率大大下降。而热红外技术有解决该问题的可能,因为人脸散发的热量可以透过面罩,被热红外相机获取。首先提出了一种面罩检测的神经网络,通过对目标检测方法进行改进,设计并实现了适用于热红外戴面罩检测的算法。实验表明,上述方法在保证检测精度的同时,具有更快的检测速度。在研究构建的数据集上,mAP为96.5%,Fps为38。其次,被面罩遮挡后,鼻子嘴巴等热可以透过面罩,并在热红外图像中有体现。利用佩戴面罩和未佩戴面罩的数据作为样本训练对抗生成网络,训练一个可以去除戴面罩样本的网络。最后提出一种九宫格损失函数,用于对去面罩后的人脸进行识别的方法。上述方法在采集86名志愿者戴面罩和未戴面罩的中波红外图像数据库中,进行了识别验证,识别率为89.68%。还提出了将来的工作方向,包括戴口罩人脸识别的研究。

With the development of artificial intelligence technology,visible light face recognition is becoming more and more popular.But visible light cannot solve the problem of facial occlusion,so it limits its application in na-tional security.Because the visible light cannot pass through the occlusion and is received by the camera,a lot of in-formation is lost.In reality,most of the features such as nose,mouth and cheek are lost,which will lead to a great de-cline in the accuracy of visible light face recognition.Thermal infrared technology can solve this problem because the heat emitted by the face can be captured by the thermal infrared camera through the mask.In this paper,a kind of neural network for mask detection is proposed.Through improving the target detection method,an algorithm suitable for thermal infrared mask detection is designed and implemented.The experimental results show that this method has a faster detection speed while ensuring the detection accuracy.When the confidence threshold is 0.9 and the mini-mum cross-union ratio is 0.85,the detection accuracy is 98.54%.Secondly,after being covered by the mask,the heat such as nose and mouth can pass through the mask and be reflected in the thermal infrared image.In this paper,the data from both wearing masks and without wearing masks are used as samples to train the confrontation generation network,and a network that can remove wearing mask samples is trained.Finally,a nine-palace loss function is pro-posed for face recognition after mask removal.This method has been applied to the MWIR image database of 86 vol-unteers with or without face masks,and the recognition rate is 89.72%.This paper also puts forward the direction of future work,including the research on face recognition with masks.