Adaptive frequency selection network for low-light image enhancement

Low-light image enhancement (LLIE) aims to enhance image brightness and contrast across diverse low-light scenarios, including uneven illumination, extreme darkness, backlighting, and nighttime conditions. Recent transformer-based LLIE advancements have demonstrated promising performance. However, existing methods face challenges in extracting and processing high-frequency information, thereby limiting their capacity to preserve fine details and mitigate artifacts. To address these challenges, we propose an adaptive frequency selection network (AFSNet) for LLIE. In this network, we introduce a frequency-selective transformer (FSFormer), which employs a novel frequency modulation strategy to adaptively adjust distinct frequency components within images. The core module of the FSFormer is the dual-branch attention module (DBAM), which involves a semantic branch and a frequency branch. The semantic branch uses the multi-dconv head transposed attention to capture contextual and overall structure information. The frequency branch incorporates a dual frequency modulation module (DFMM) to achieve region-adaptive frequency decomposition and enhancement. Additionally, a frequency-aware feature fusion module (Freqfusion) facilitates feature fusion across different scales. Extensive experimental results demonstrate that our approach achieves competitive performance across multiple synthetic and real-world datasets compared to state-of-the-art (SOTA) methods.