基于红外热图与深度学习的建筑室内人脸属性分类研究

(同济大学 土木工程学院,上海 200092)

热舒适; 计算机视觉; 红外热成像; 年龄识别; 性别识别

Deep learning-based facial attribute classification from indoor thermal images
LI Peixian, CAO Daqian, DAI Pengfei, LU Yujie, LIU Bo

(College of Civil Engineering, Tongji University, Shanghai 200092, China)

thermal comfort; computer vision; thermography camera; age classification; gender recognition

DOI: 10.15986/j.1006-7930.2022.03.015

备注

利用热成像相机预测个体热舒适是无干扰温控的一种途径,有助于建筑节能.而人体热舒适范围在不同年龄与性别之间差异较大,现有文献尚缺乏红外热图中年龄与性别差异的研究.为探究利用深度学习从红外热图中自动识别性别与年龄的可行性,本文建立了红外热图和可见光的人脸数据集,对比了ResNet-50、DenseNet-121、DenseNet-201、Inception-V3四种卷积神经网络的效果,实验结果表明:男女红外热图差异明显,用Inception-V3可达到98.7%的识别准确率; 中青年红外热图差异较小,中老年红外热图差异明显,在分三类时,ResNet-50可获得80.0%的年龄识别准确率; 性别与年龄识别准确率均高于现有文献记载.同时,本文研究了红外滤镜和人脸裁剪对准确率的影响,提出了有助于提高识别精度的人脸红外热图数据采集与处理方法.
Thermal camera is a non-invasive method to predict individual thermal comfort which helps saving HVAC energy. While thermal comfort differs a lot between different groups of age and gender, existing literature lack research on the detection of age and gender using thermal cameras for accurate prediction of human thermal comfort in the built environment. To explore the feasibility of using deep learning to automatically recognize gender and age from thermal images, we establish a dataset of thermal images and visible-light images, study the impacts of algorithms(four convolutional neural networks: ResNet-50, DenseNet-121, DenseNet-201, and Inception-V3), thermal image filters, and image cropping on the recognition accuracy, and compare the recognition performances using thermal images and visible-light images. The results show that the gender classification accuracy can reach 98.7% using Inception-V3, meaning that there is a significant difference between male and female thermal images. The highest age classification accuracy(80.0%)is achieved using ResNet-50 when the dataset is divided into three classes—young, middle-aged, and old. It is noticed that there is little difference between young and middle-aged thermal images but a more obvious difference between the middle-aged and the old ones. The achieved accuracies are higher than or comparable to those in the literature. This study demonstrates that convolutional neural network is suitable for gender and age recognition from thermal images.