SNOC: Subtle Nested Objective Configuration for Joint Ultra-Low-Light Enhancement and Super-Resolution

Ultra-low-light image restoration remains challenging in part because static optimization objectives are often brittle across the diverse degradation patterns induced by extreme darkness. Existing joint enhancement and super-resolution paradigms predominantly rely on manually specified objective trade-offs, leading to scene-dependent failures such as color bias, exposure inconsistency, and artifact propagation. SNOC integrates a subtle rectification architecture with adaptive objective-portfolio configuration. Architecturally, it employs Semantic Illumination Cross Calibration (SICC), Exposure-aware Rectification Unit (EaRU), and Grid-aware Dynamic Up-sampler (GaDU) to coordinate illumination-aware, exposure-aware, and detail-oriented representations. Beyond architectural design, it introduces a compact nested objective configuration mechanism that adaptively updates learnable coefficients over a comprehensive objective portfolio, aligning module-specific restoration targets with global image quality and avoiding labor-intensive manual objective tuning.

Recommended citation: Jiaxin Gao, Yaohua Liu*, Danchen Cui, Zhihui Zhao. SNOC: Subtle Nested Objective Configuration for Joint Ultra-Low-Light Enhancement and Super-Resolution[C]. ACM International Conference on Multimedia Retrieval (ICMR), 2026.
Download Paper