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

TL;DR — TCA-Net is a Low-Light Image Enhancement network that improves how intensity and chromaticity streams are fused in HVI space using a confidence-thresholded cross-attention, mattering because reliable fusion largely determines final image quality (noise, color fidelity, efficiency).

• Replaces fixed-quota Top-K sparse attention with a fixed confidence threshold whose retained cardinality is input- and layer-adaptive, keeping high-confidence cross-stream interactions and suppressing unreliable ones—motivated by the observation that cross-attention confidence is strongly layer-dependent.
• Adds a Phase-guided Fourier Interaction Module for structure-aware brightness initialization before fusion, and a Decoupled Dual-Stream Guidance Module that builds residual intensity features to suppress chromaticity leakage during reconstruction.
• Uses Scale-Aware Consistency Regularization during training for structural robustness under scale perturbations.
• Reported as competitive on LOL-v1, LOL-v2, Sony-Total-Dark, and LSRW-Huawei with better color fidelity and a compact parameter size (specific metrics not provided in the abstract).