A Soft Calibration Technique for Infrared Imaging based on Physical Model of Thermal Radiation and Variational Bayesian Inference from Small Dataset

Nowadays, infrared thermal imaging has been widely used as a non-contact, long-distance, and large-scale temperature measurement method. Nonetheless, the temperature measurement process of the infrared sensor is affected by several factors, such as the measurement distance, the emissivity of the object’s surface, the ambient temperature, and so on. Therefore, there is still considerable error between the measured temperature and the ground-truth one. Two significant factors severely limit the temperature calibration: the complex factors mentioned above and the lack of big data for learning nonlinear-mapping. This paper proposes a soft-calibration approach of infrared imaging temperature based upon the combination of the physical model of thermal radiation (PMTR) and variational Bayesian inference (VBI). Rather than big-data mining, the proposed method can combine PMTR and the sparsity-enforcing prior of the infrared image, therefore using the VBI algorithm to learn the nonlinear mapping relationships from a limited dataset. Moreover, the exponential potential prior is properly assigned to the Bayesian inference process and the Kullback-Leibler (KL) divergence criterion is adopted for the optimization process. The soft calibration of a cost-effective of infrared sensor has been carried out through the experiment. The result shows that the temperature calibration errors of our method are within $1.0 ^{\circ}\mathrm{C}$, which is better than the Maximum A Posterior (MAP) method, the artificial neural network (ANN) method, and so on.