Rethinking progressive low-light image enhancement: A frequency-aware tripartite multi-scale network

Low-light images often suffer from complex degradation factors such as reduced contrast, color distortion, and noise. Recently, progressive networks have gained great popularity in low-light image enhancement due to their strong ability to recover image details and improve overall quality. However, most existing progressive methods typically rely on a single-scale fusion strategy, which limits their capacity to capture rich multi-scale information effectively. To overcome this limitation, we propose a novel Progressive Tripartite Multi-Scale Network (PTMSNet) for low-light image enhancement. Our network uniquely enhances multi-scale interactions among different branches across multiple scales to achieve comprehensive feature fusion. To be more specific, we design a Parallel Hybrid Module (PHM) that integrates a Transformer branch with a Convolutional Neural Network (CNN) branch in parallel. This design allows the two branches to learn complementary feature representations and effectively preserve local textures and structural information. Meanwhile, we introduce a Frequency-Aware Fusion Module (FAFM), which leverages global frequency-domain information to generate a more accurate Signal-to-Noise Ratio (SNR) map than previous Gaussian-based SNR fusion. A joint loss function combining pixel, structural, and color constraints is further employed to optimize model parameters. Extensive evaluations on three benchmark datasets demonstrate that PTMSNet significantly outperforms state-of-the-art (SOTA) competitors, including several progressive predecessors. Moreover, its strong performance on the ExDark dataset and in real-world scenario generalization experiments highlights robust low-light detection capability and practical applicability in photometric measurement systems with unstable illumination. The code will be made publicly available at https://github.com/kbzhang0505/PTMSNet.