基于倒残差多尺度卷积注意力的红外热成像人脸对齐算法  

Face Alignment in Thermal Infrared Images Based on Inverted Residual Multi-scale Convolutional Attention

作　　者：刘旭龙 李枭 许爽 贾紫巍 LIU Xulong;LI Xiao;XU Shuang;JIA Ziwei(School of Computer and Communication Engineering,Northeastern University at Qinhuangdao,Qinhuangdao,Hebei 066000,China)

机构地区：[1]东北大学秦皇岛分校计算机与通信工程学院,河北秦皇岛066000

出　　处：《计量学报》2024年第11期1634-1641,共8页Acta Metrologica Sinica

基　　金：国家自然科学基金(61401080);河北省自然科学基金(F2019501101)。

摘　　要：红外热成像人脸对齐在提取人脸面部温度方面起着至关重要的作用,其定位精度直接关系到对人脸各区域温度数据的准确采集与分析。当前,主要的人脸对齐方法多数应用于可见光人脸图像,但将其直接应用于红外热人脸图像存在精度不足等问题。为解决这一问题,提出了一种基于多尺度卷积注意力的热红外人脸对齐算法。利用多尺度信息的优势,将多尺度卷积注意力机制与倒置残差卷积网络相结合,并引入wing loss作为损失函数,增强了网络模型的特征提取能力。实验结果显示,所提出的算法在公开热红外人脸数据集和自采的面瘫人脸数据集上的归一化平均误差分别达到了3.23%和3.94%。相较于传统方法,该算法针对各器官的定位精度均有一定程度的提高,应用范围可扩大至面瘫人群,意义较大。Face alignment in thermal infrared images is crucial for accurately extracting facial temperature data,as its positioning precision directly impacts the accuracy of temperature analysis in various facial regions.However,most prevalent face alignment algorithms designed for visible face images encounter limitations when directly applied to infrared thermal images,resulting in insufficient accuracy.To address this issue,a face alignment algorithm specifically designed for thermal infrared images,leveraging a multi-scale convolution attention mechanism,is introduced.This algorithm effectively integrates the multi-scale convolutional attention mechanism with an inverted residual convolutional network,while incorporating the wing loss as the loss function to further enhance the network model's feature extraction capabilities.On both an open thermal infrared face dataset and a self-collected facial palsy dataset,the algorithm achieves normalized mean errors of 3.23%and 3.94%,respectively.This represents a significant improvement in localization accuracy for facial features compared to traditional methods,extending its applicability to populations with facial palsy.This advancement holds immense potential for various applications.

关 键 词：图像处理 红外热成像 人脸对齐 多尺度卷积注意力 归一化平均误差 

分 类 号：TB96[机械工程—光学工程]