Estimating temperatures with low-cost infrared cameras using physically-constrained deep neural networks.

Area of Science:

• * Thermal imaging
• * Computer vision
• * Sensor calibration

Background:

• * Low-cost thermal cameras exhibit significant inaccuracies (±3°C) and spatial nonuniformity.
• * These errors are dependent on the camera's ambient temperature.
• * Accurate temperature estimation and nonuniformity correction are crucial for reliable thermal imaging.

Purpose of the Study:

• * To develop a method for accurate temperature estimation using low-cost infrared cameras.
• * To rectify space-variant nonuniformity in thermal camera detectors.
• * To improve the reliability and accuracy of thermal imaging systems.

Main Methods:

• * Development of a nonuniformity simulator accounting for ambient temperature.
• * Introduction of an end-to-end neural network integrating a physical camera model and ambient temperature.
• * Training the neural network with simulated data for temperature estimation and nonuniformity correction from single images.

Main Results:

• * Significant reduction in mean temperature error compared to state-of-the-art methods (0.29°C gap).
• * Further error reduction of 0.1°C achieved by constraining the physical model within the network.
• * Mean temperature error of 0.37°C over an extensive validation dataset.
• * Successful verification on real-world field data with equivalent results.

Conclusions:

• * The proposed neural network method effectively corrects thermal camera nonuniformity and estimates temperature accurately.
• * The integration of a physical camera model enhances the precision of temperature measurements.
• * This approach offers a low-cost, high-accuracy solution for thermal imaging applications.