Thermal imager performance assessment: MTDP prediction from camera images

The minimum temperature difference perceived (MTDP) is a well-known performance test of thermal imagers. The measurement is based on the perception of a four-bar target by a human observer and applies to well and under sampled imagers. The MTDP is a central quantity of the range performance model TRM4 and an analytical expression exists to calculate MTDP from known imager specifications. In this work, we investigate regression models for the prediction of MTDP from images of degraded four-bar targets. Such models are interesting from two perspectives. First, they might be a step towards an automated MTDP measurement and, second, they can be used to assess imagers with embedded digital signal processing in contrast to analytical MTDP calculation. The degraded images are simulated using the image simulation software OSIS from Fraunhofer IOSB, which simulates camera effects such as optical diffraction and temporal and spatial noise based on TRM4 specifications. Several camera properties and scene settings are varied to generate datasets for model training and validation. For each configuration the corresponding MTDP values are calculated using TRM4. Then CNN-based regression models are trained to predict the MTDP values from the simulated camera images. Dependencies of model performance on target size and individual image degradations simulated by OSIS and the incorporation of TRM4 camera specifications into the model are presented. The potential for noise reduction is investigated by averaging model predictions over image stacks, each based on a unique camera specification. Furthermore, the inclusion of recorded image data and measured MTDP values in the training process is discussed.