Thermal image restoration that learns and forgets continuously

Thermal infrared imaging degrades in ways that change over time—rain, fog, noise—but existing restoration methods assume a fixed set of degradations. ECMRNet addresses this by framing thermal image restoration as a continual learning problem. The network expands with new degradation types, compresses itself via Structural Entropy Pruning to remove redundant parameters, and mines historical representations to handle compound degradations. Experiments show it outperforms all-in-one baselines on both single and compound degradations while using fewer parameters. Code is released.