基于面部表情识别的室内热声环境评价研究
Research on indoor thermoacoustic environment evaluation based on facial expression recognition
摘要：
基于图像识别、面部特征提取和机器学习算法,构建了基于人员面部表情识别的室内环境舒适度评价模型。在人工气候室采集了24名受试者的面部表情,使用残差神经网络通过自动提取特征并进行学习,实现了87.7%的验证集准确率。建立了舒适、中性、不舒适3种表情数据库,通过学习不同整体舒适度下表情图像的视觉特征,使训练好的模型能够通过人员面部表情识别对人员的整体舒适状态进行评价。随后,采用方向梯度直方图算法提取面部特征,对12种不同机器学习算法的准确率、训练时间和泛化能力进行了比较,发现K-近邻算法的准确率可达83.2%,且用时较短。因此,本文构建的基于表情识别的室内环境舒适性评价模型可为热声环境多因素作用下人体舒适性感知的相关研究提供科学参考。
Abstract：
This study constructs an indoor environment comfort evaluation model based on human facial expression recognition by using image recognition, facial feature extraction, and machine learning algorithms. The facial expressions of 24 subjects are collected in an artificial climate chamber and the residual neural network (ResNet) is used to automatically extract features and conduct learning, resulting in a high validation set accuracy of 87.7%. Three expression databases of comfort level: comfortable, neutral, and uncomfortable are established. By learning the visual features of expression images under different overall comfort levels, the trained model can hence evaluate the comfort levels of personnel through facial expression recognition. Subsequently, the histogram of oriented gradients (HOG) algorithm is used to extract facial features, and the accuracy, training time, and generalization ability of 12 different machine learning algorithms are compared. It is found that the K-nearest neighbor (KNN) algorithm can reach the highest accuracy of 83.2% with less time. Consequently, the indoor environment comfort evaluation model based on facial expression recognition constructed in this paper can provide a scientific reference for relevant research on human comfort perception under the combined action of multiple factors in different thermoacoustic environments.
Keywords:indoor thermoacoustic environment; facial expression; image recognition; comfort evaluation model; machine learning