使用高分辨率网络在热红外图像上提取人脸关键区域温度

Applying the high-resolution network technology to extract face area temperature from thermal images

近年来,个人热舒适模型取得了众多学者的研究和关注。其中,脸部皮肤表面关键区域的温度是模型中的关键输入特征。使用高分辨率特征学习网络(HRNet)在热红外图像上直接进行人脸关键点检测,以提取人脸关键区域温度,其准确率高于同类热红外人脸关键点识别算法的最佳值。还将人脸属性作为权重加入到损失函数中,进一步提升了对于难例(如侧脸、遮挡、夸张表情)的识别准确率。最后,对比了热红外图像与不同环境下可见光图像的检测效果,证明了在一些特殊条件下(如低光照、隐私保护),热红外图像可以代替可见光图像进行准确的关键点识别。

In recent years,personal comfort model has attracted much attention.The temperature of facial skin surface is the key parameter in the model.This paper used high-resolution network(HRNet)to directly detect facial landmark on thermal images to extract facial temperature.Achieve the superior accuracy over the state-of-the-art algorithm.Furthermore,we add face attributes as weights to the loss function to further improve the accuracy of difficult cases(such as profile faces,occlusions,exaggerated expressions).Finally,compare its performance on infrared images and RGB images in different conditions,and prove that infrared images can replace visible light images for accurate landmark detection under some special conditions(such as low light,privacy protection).