HFCNet: A hybrid CNN-transformer network with frequency-aware gating and cross-stream interaction for low-light image enhancement

Low-light image enhancement (LLIE) aims to restore the visibility and details of images captured under extreme photon-starved conditions. However, when dealing with extreme or spatially varying illumination, existing methods frequently struggle with severe color distortion, the entanglement of high-frequency detail loss and noise amplification, as well as local exposure inconsistencies. In response to these significant obstacles, we introduce HFCNet, a hybrid CNN-Transformer network empowered by frequency-aware gating and cross-stream interaction mechanisms. First, to break the deep coupling between luminance and chrominance, the model performs dual-stream feature extraction within the polarized HVI color space. An Interactive Fusion Block (IFB) is accordingly designed to dynamically constrain color boundaries using structural priors, reducing color shifts and semantic fragmentation. Second, to mitigate noise contamination during feature scale transformations, the network employs the Discrete Wavelet Transform (DWT) for faithful downsampling. Furthermore, a Learnable High-Frequency Gating (LHFG) mechanism is innovatively introduced in the decoding stage to adaptively filter out shot noise from skip connections while purifying essential physical textures. Finally, tackling the challenge of spatially varying illumination, we construct a Dual-Domain Block (DDB) for spatial-frequency synergy and integrate a Transformer bottleneck layer. This combination modulates global contrast in the frequency domain and captures long-range illumination dependencies, enhancing the global consistency of scene brightness. Extensive experiments on the LOL-series paired datasets and five no-reference real-world datasets demonstrate that HFCNet achieves competitive performance against existing state-of-the-art (SOTA) methods in both quantitative metrics and visual perceptual quality, showing robust capabilities in detail preservation, highlight suppression, and color fidelity.