SUPER-RESOLUTION RECONSTRUCTION AND HIGH-PRECISION TEMPERATURE MEASUREMENT OF THERMAL IMAGES UNDER HIGH- TEMPERATURE SCENES BASED ON NEURAL NETWORK

Abstract

Accurate temperature readings are vital in fire resistance tests, but conventional thermal imagers often lack sufficient resolution, and applying super-resolution algorithms can disrupt the temperature and color correspondence, leading to limited efficiency. To address these issues, a convolutional network tailored for high-temperature scenes is designed for image super-resolution with the internal joint attention sub-residual blocks (JASRB) efficiently integrating channel, spatial attention mechanisms, and convolutional modules. Furthermore, a segmented method is developed for predicting thermal image temperature using color temperature measurements and an interpretable artificial neural network. This approach predicts temperatures in super-resolution thermal images ranging from 400 to 1200°C. Through comparative validation, it is found that the three-neuron neural network approach demonstrates superior prediction accuracy compared to other machine learning methods. The seamlessly combined proposed super-resolution architecture with the temperature measurement method has a predicted RMSE of 20°C for the whole temperature range with over 85% of samples falling within errors of 30°C.