Temperature Measurement of Metal Surface at Normal Temperatures by Visible Images and Machine Learning

Thermal radiant energy of metal surfaces is weak in visible bands at normal temperatures, making it difficult to perform non-contact temperature measurements by visible images based on thermal radiation principle. However, this paper proposed an intelligent temperature measurement method of metal surfaces at normal temperatures in sunlight based on thermal-modulated reflection. A digital camera was used to take photos of iron, aluminum alloy, and copper, with their temperature ranging from 26.0°C to 100.0°C. These images composed three corresponding image libraries of these materials. For each image in each library, two kinds of statistical features, RGB gray level histograms (RGB-GLHs) and deep semantic chromatic features (DSCFs), were extracted and labeled by the image’s corresponding measured temperature, forming two kinds of Feature-Label datasets of the image library. For each library, both kinds of Feature-Label datasets were used to train machine learning (ML) models for temperature prediction. Besides, a baseline model, Resnet50, was trained for temperature prediction. Results showed that the trained ML models predicted the surface temperature of these materials well. Models trained by DSCFs greatly improved prediction accuracy compared with those trained by RGB-GLHs and Resnet50. The K-Nearest Neighbor models had a mean absolute error under 1.0. Meanwhile, using DSCFs significantly saved the calculation time and data storage spaces. This method would provide a new choice for the temperature measurement of electrical equipment’s metal parts and other cases requiring non-contact measurements.