Thermal imaging system performance is described by the minimum resolvable temperature difference (MRTD), which is the spatial-frequency-dependent thermal sensitivity. Three different MRTD models are presented. Two are based on the adaptive matched filter concept in order to describe the human eye-brain combination. The third uses the limited synchronous integrator concept. In all three models the usual assumptions of periodicity and of one-dimensional targets are replaced by two-dimensional and more realistic representations of the targets. MRTD expressions are derived for these three models, and the values calculated from them are compared with the Ratches-Lawson model predictions and with measured results for two different systems with known parameters. It is shown that the limited synchronous integrator model results in a more accurate functional dependence of the MRTD on the spatial frequency. The prediction error commonly found in the low frequency region is reduced, and the overall accuracy is enhanced throughout the frequency range of the system in comparison to the Ratches-Lawson model.