Corrosion in metals is simulated with variations in plate thickness which are laterally infinite in a 1D model and are represented by milled flat-bottom holes in a 2D model. Temperature contrast over corroded areas is chosen as an informative parameter, quite independent of absorbed thermal energy in the infra-red thermographic test. It is shown that, due to lower sensitivity to rear-side effects at the beginning of the thermal process-and increasing 3D hear diffusion effects at the end of the process, there is an optimum time to detect corrosion. A robust inversion function is proposed and its stability against variations in tested material, heat pulse duration and observation time is analysed using numerical modelling. Corrosion in a steel specimen of 1.3 mm thickness is experimentally studied, having proved the validity of the inversion algorithm with an average accuracy of 17% for material loss ranging from 74 to 14%.