An Energy-Domain IR NUC Method Based on Unsupervised Learning

To obtain accurate blackbody temperature, emissivity, and waveband measurements, an energy-domain infrared nonuniformity method based on unsupervised learning is proposed. This method exploits the inherent physical correlation within the calibration dataset and sets the average predicted energy-domain value of the same blackbody temperature as the learning goal. Then, the coefficients of the model are learned without theoretical radiance labels by leveraging clustering-based unsupervised learning methodologies. Finally, several experiments are performed on a mid-wave infrared system. The results show that the trained correction network is uniform and produces stable outputs when the integration time and attenuator change within the optimal dynamic range. The maximum change in the image corrected using the proposed algorithm was 1.29%.