Correction of emissivity in thermograms with neural networks

Infrared thermography, a non-destructive testing technique, enables the measurement and visualisation of surface temperatures by capturing images of emitted infrared radiation. To achieve accurate results, it is essential to correct for emissivity, a material-dependent property that affects the amount of emitted infrared radiation and distorts apparent temperatures across different materials. This study proposes a new approach for correcting thermograms by segmenting an image of the same object captured in the visible light range into material classes using a convolutional neural network. For each material class, the emissivity is determined and used for the correction. The methodology includes setting up experiments for image acquisition, determining emissivity through controlled experiments, and registering the infrared images and visible light images. Unlike previous methods, this approach can be designed to be fully automated in the future and is robust to heterogeneous temperature distributions. Two experiments are conducted in this study: a proof of concept study involving specimens made of different materials and a study focusing on a printed circuit board. It was found that the correction procedure worked for both case studies. However, the quality of the correction is highly depended on the accuracy of the individual steps in the correction procedure.