Application of visible light polarization imaging in normal temperature measurement

An approach to measure the temperature of objects in a common range (30 ~ 150 °C) using visible light polarization imaging and Convolutional Neural Networks (CNN) is proposed. Objects at different temperatures is imaged by visible light polarization imaging to obtain images of intensity distribution, images of angle of polarization (AOP) distribution and images of degree of polarization (DOP) distribution, which are analyzed by the CNN to determine the objects temperature. The experimental results show that the CNN can derive the objects temperature based on the intensity image, AOP image and DOP image obtained from the reflected light from the objects surface. It indicates that not only the reflectivity of the light reflected from the target surface changes with the objects temperature, but also its polarization characteristics. By comparing the temperature detection error of the CNN for the three types of images, it was found that temperature detection using DOP images has better accuracy, whose mean absolute error is 0.5 °C lower than that of detection using intensity images. Moreover, temperature detection based on the CNN and DOP image shows greater robustness to background illumination variations. This approach is expected to be widely used in the application of measuring objects temperature of common range based on visible images.