This project investigates the design of dedicated neural architectures for industrial imaging inverse problems by incorporating modality-specific image formation models and noise characteristics directly into the network structure through algorithm unrolling. By unfolding iterative optimization algorithms into learnable layers, the resulting architectures encode domain knowledge of the forward model, enabling principled regularization and reducing dependence on large training datasets. The approach targets multiple imaging modalities including infrared, multi-spectral, and 3D geometry under challenging conditions such as low light, motion blur, and occlusions, with expected gains of 2–5 dB in visual quality and a fivefold reduction in model parameters, making the architectures suitable for deployment on resource-constrained edge and IoT nodes.