SOGKD-Lite: Lightweight Semantic Object-Guided Network for Efficient Low-Light Image Enhancement via Knowledge Distillation

Low-light image enhancement techniques have advanced significantly, yet existing models often face challenges with deployment efficiency due to their high computational complexity. To bridge this gap, we propose SOGKD-Lite, a lightweight semantic object-guided network leveraging knowledge distillation for efficient low-light image enhancement. Our method systematically reduces the computational complexity of the CIDNet architecture by shrinking its feature channel width, resulting in a 38% reduction in both parameters and FLOPs. To mitigate the performance loss typically associated with lightweight models, we introduce a multi-guided loss suite that includes feature-domain alignment via knowledge distillation, frequency-domain constraints through a Wavelet loss, and a novel semantic-domain color consistency loss. This semantic loss uniquely employs the Segment Anything Model 2 (SAM 2) to maintain region-aware color fidelity between enhanced images and ground truth. Extensive experiments on LOLv1, LOLv2, SID, and LOL-Blur datasets demonstrate that SOGKD-Lite achieves competitive performance against state-of-the-art methods, establishing a new benchmark in balancing efficiency and performance for low-light image enhancement.