DiLien: Domain-Incremental Low-Light Image Enhancement

In this work, we introduce a novel domain-incremental learning framework for low-light image enhancement. Unlike most existing methods tailored to single scenes, our framework adapts to dynamically evolving environments by revisiting old domain knowledge and efficiently acquiring new domain knowledge without retaining any old samples. It comprises two key components: feature-guided adaptation and a domain transfer-and-stable discriminator. The feature-guided adaptation leverages intermediate feature representations of new-domain inputs within the old stage network to transfer old domain semantic knowledge through distillation. Meanwhile, the domain transfer-and-stable discriminator employs an adversarial strategy, enabling the network to learn from the old stage network’s outputs while preventing over-learning of old domain information. Additionally, we collect an incremental dataset specifically for low-light image enhancement, and extensive experiments validate the effectiveness of our method.