Thresholded Cross-Attention for Reliable Intensity-Chromaticity Fusion in Low-Light Image Enhancement

Low-light image enhancement (LLIE) often struggles with the entanglement of brightness, color, and noise. While recent methods use HVI color space decomposition to separate intensity and chromaticity, the subsequent fusion of these streams remains problematic. The authors investigate why dense or fixed-quota (Top-K) cross-attention mechanisms fail to produce reliable fusion, specifically noting that attention confidence varies significantly across different network layers and input conditions.

The authors propose TCA-Net, a network centered on Thresholded Cross-Attention (TCA). Unlike traditional Top-K attention, which forces a fixed number of interactions regardless of their reliability, TCA uses a fixed confidence threshold. This allows the network to dynamically retain only high-confidence cross-stream dependencies while discarding unreliable, noise-correlated ones.

To support this core mechanism, the authors introduce two auxiliary modules:

• Phase-guided Fourier Interaction Module (PFIM): This module processes intensity features in the frequency domain, using phase information to provide a structure-aware brightness initialization before fusion.
• Decoupled Dual-Stream Guidance Module (DDSGM): This module constructs residual intensity features to suppress chromaticity leakage, ensuring that color restoration remains stable during the final reconstruction.

Additionally, the authors employ Scale-Aware Consistency Regularization (SACR) during training to improve the model's robustness against scale perturbations.

Experiments across several benchmarks (LOL-v1, LOL-v2, Sony-Total-Dark, and LSRW-Huawei) demonstrate that TCA-Net achieves competitive restoration accuracy and superior color fidelity compared to existing state-of-the-art methods. The authors show that the threshold-based approach effectively adapts to layer-dependent attention distributions, avoiding the systematic errors introduced by fixed-quota pruning. The resulting architecture is also noted for its compact parameter size, making it efficient for practical deployment.