Mean‐Local Binary Pattern‐Guided Multi‐Attention Network for Low‐Light Image Enhancement

ABSTRACT Low‐light image enhancement faces key challenges: noise is hard to remove and may be amplified, while residual dark regions, artefacts and detail loss often remain post‐enhancement. To address the difficulty of simultaneous denoising and enhancement, this paper proposes a mean‐local binary pattern (M‐LBP)‐guided multi‐attention network (MGA‐LLIEN) for low‐light image enhancement. Leveraging local binary patterns’ (LBP) high sensitivity to texture and illumination, the network enables adaptive brightness adjustment, mitigates detail loss and noise amplification and facilitates detail recovery. The network operates in four steps: (1) A cyclic decomposition network splits low‐light images into independent illumination and reflection components. (2) An improved M‐LBP feature extraction method is proposed, using M‐LBP to guide illumination component enhancement and reflection component denoising. (3) A multi‐level channel attention mechanism is integrated into the illumination enhancement sub‐network to ensure colour recovery aligned with real scenarios. (4) A dual‐branch denoising module processes reflection components: the upper branch inverts images to enhance low‐contrast region information, excavate noise‐obscured details and suppress high‐frequency noise. Finally, processed components are recombined for enhanced output. Ablation experiments validate each module's rationality. Extensive tests on public datasets show MGA‐LLIEN outperforms state‐of‐the‐art methods in subjective visual quality and objective metrics, confirming its superiority.