MFGDAFormer: Multi-scale frequency-guided dual-branch attention transformer for low-light image enhancement

Images captured in low-light conditions, such as at night or in backlit environments, often suffer from uneven illumination, insufficient contrast, and significant noise, which compromise the accuracy and robustness of visual tasks like object detection. Low-light image enhancement techniques are essential for improving visibility, restoring details, and reducing noise, thereby enhancing the performance of various visual tasks. Effective illumination adjustment and detail preservation are key to achieving successful low-light image enhancement. However, existing methods face significant challenges, including inadequate modeling of lighting patterns, insufficient local feature representation, neglect of frequency-domain information, and difficulty balancing noise suppression with detail preservation. This paper introduces the Multi-scale Frequency-Guided Dual-branch Attention Transformer (MFGDAFormer) for low-light image enhancement. The framework incorporates a CNN-based illumination estimator to guide enhancement and lighting adjustment, while the Fourier Sparse Multi-scale Attention Mechanism (FSMAM) facilitates effective frequency-domain analysis, adaptive feature modulation, and multi-scale detail preservation. Furthermore, the U-Net architecture is employed for fine-grained illumination enhancement, and the Dual-Large Kernel Activation Attention Module (DLKA) integrates large-kernel convolutions with adaptive attention mechanisms to optimize feature fusion. Experimental results demonstrate that MFGDAFormer outperforms state-of-the-art models, with PSNR improvements ranging from 0.21 dB to 0.88 dB across multiple low-light datasets.