Model-Based Neural Networks for Thermographic Image Reconstruction

Thermographic imaging is a specific non-destructive evaluation (NDE) approach in which the investigated object is exposed to an initial thermal excitation. The utilization of in-frared cameras to detect the induced surface temperature change enables the deduction of the internal structure of the inspected material. This requires addressing a large-scale ill-posed linear inverse problem, involving regularization and forward modeling of the observed thermal diffusion process. In this paper, we introduce a hybrid approach for ther-mographic imaging, which combines a deep unfolded Fast It-erative Shrinkage-Thresholding Algorithm (FISTA) algorithm with a U-Net. Consequently, the resulting network architecture incorporates established numerical regularization heuristics, such as sparsity and smoothness. Additionally, to mitigate the data demand of DL approaches, we establish a simulated training database containing both point-like defects and line-like cracks. As a case study, we begin by considering applications in non-destructive material testing, starting with 2D problems and progressing towards thermal tomography in 3D.