TCA-Net Rethinks Intensity-Chromaticity Fusion for Low-Light Image Enhancement

Introduction

Low-light image enhancement is often described as the task of making dark images brighter, but the real challenge is more delicate: a useful model must lift illumination while avoiding noise amplification, color distortion, and excessive computational cost. Recent HVI-based approaches try to reduce color entanglement by separating intensity from chromaticity. The arXiv paper “Thresholded Cross-Attention for Reliable Intensity-Chromaticity Fusion in Low-Light Image Enhancement” focuses on what happens after that separation: how the two streams are fused again.

The authors argue that this fusion step is an underexplored factor behind final image quality. If the network reconnects intensity and chromaticity features in an unreliable way, the benefits of decoupling may be weakened by noisy dependencies or color leakage during reconstruction.

Key ideas

• A limitation of fixed Top-K attention. Sparse attention mechanisms often retain a fixed number of strongest interactions. According to the paper, cross-stream attention confidence is strongly layer-dependent. A fixed quota can therefore remove useful dependencies in some layers while preserving noisy ones in others.

• Thresholded Cross-Attention as the core. TCA-Net replaces the rigid Top-K selection rule with a fixed confidence threshold. Instead of keeping the same number of interactions everywhere, the number of retained cross-stream links becomes adaptive to both the input and the layer. Only interactions above the confidence threshold are kept.

• Working within HVI rather than inventing a new color space. The method does not present another color representation as its main contribution. It instead improves the reliability of fusion in the HVI space, where intensity and chromaticity have already been separated.

• Cleaning the signal before and after fusion. Before fusion, a Phase-guided Fourier Interaction Module provides a structure-aware brightness initialization for the intensity stream. Afterward, a Decoupled Dual-Stream Guidance Module builds residual intensity features to reduce chromaticity leakage during reconstruction.

• Scale-aware training regularization. The proposed Scale-Aware Consistency Regularization is designed to improve structural robustness when images undergo scale perturbations during training.

Why it matters

The interesting part of TCA-Net is not simply that it adds another attention module to low-light enhancement. Its contribution is more specific: it questions whether a fixed sparse-attention budget is appropriate when the reliability of intensity-chromaticity dependencies changes across layers and inputs. A confidence threshold is a more selective mechanism, allowing dense retention when many interactions are trustworthy and stronger pruning when they are not.

According to the paper abstract, experiments on LOL-v1, LOL-v2, Sony-Total-Dark, and LSRW-Huawei show competitive restoration accuracy, improved color fidelity, and a compact parameter size. For real-world LLIE systems, that combination is important because enhancement quality, color stability, noise control, and efficiency all matter at once.

The current arXiv material does not provide enough space in the abstract to judge every detail, such as exact metric gains or failure cases. Still, the paper highlights a useful direction: low-light enhancement may improve not only through better decomposition, but also through more trustworthy recomposition of brightness and color information.

Source: arXiv