High-precision thermography based on JMAP inference for human face temperature

During the pandemic Covid-19, infrared thermography is an efficient way to detect susceptible persons with abnormal temperatures. However, two significant factors seriously limit the temperature precision: measurement uncertainty of the infrared camera and model inaccuracy of thermal radiation. In this paper, we propose the joint maximum a posterior (JMAP) approach with a new hierarchical prior model. The advantages of JMAP are that the Bayesian inference can combine prior model and likelihood model to regulate the uncertainty from both physics and measurements. At first, we obtain the estimated parameters of the thermal radiation model from training-data. We propose that the distribution of actual temperature and the distribution of measurement error satisfy the Gaussian distribution. We take the variance of the Gaussian distribution as the latent variable and assume that the variance satisfies the inverse gamma distribution, which we control by setting hyperparameters, which determine the uncertainty of the temperature variance so that variances are updated continuously rather than constants. We apply the JMAP inference to test human-face temperature for infrared camera calibration. Although the pre-set parameters of the cost-effective infrared camera badly affect the accuracy, our proposed approach can refine the error of less than 0.1. compared with 1.0. without calibration at certain distances. Therefore, our proposed approach can offer an efficient and accurate way to screen people with abnormal body temperature against Covid-19.