We present a novel multispectral imaging system that measures temperature without knowledge of emissivity. It combines support vector machine regression with low cost PbSe imagers, sensitive in the MWIR range and capable to achieve very high speed acquisition rates with a medium resolution. The system is modular and builds on two or more apertures sensitive to different but close spectral bands. Inspired by the approach adopted by ratio and multi wavelength pyrometers, we estimate temperature from the combined response at these bands. However, we adopt a flexible and transparent approach to modeling multiple regression based on machine learning and using synthetic datasets. We demonstrate high accuracy and robustness against variations in the value of emissivity. Besides a working prototype, our contribution renders a simple procedure for the design of cost effective thermographic systems for field applications demanding reliable measurements in unconstrained conditions.

Back