SFANet: Spatial-Frequency Aggregate Network for Multi-Exposure Image Fusion

Multi-Exposure image Fusion (MEF) aims to combine the complementary information in the over-underexposed source image pairs or sequences to obtain a fused image with rich texture details, appropriate brightness, and pleasure visual quality. Existing multi-exposure image fusion networks predominantly employ Convolutional Neural Networks (CNN), which exhibit limited capacity in capturing global context information. For this reason, we design a network based on spatial-frequency domain aggregate, named SFANet. In general, the network adopts a U-shape structure. Firstly, we extract different scale of feature maps via down-sampling. Then Spatial-Frequency Aggregation Block (SFAB) is proposed for feature fusion. Specifically, this block consists of three modules: Spatial Fusion Module (SFM), Frequency Fusion Module (FFM), and Spatial-Frequency Fusion Module (SFFM). Specially, SFM is designed to integrate spatial features across different scales and extracts local information from source images, while the FFM leverages the characteristics of the Fast Fourier Transform (FFT) to capture long-term dependencies in the source image. To enhance the interaction between the spatial-frequency domain and to capture complementary information from the source image, we suggest SFFMs. Extensive evaluations show that the proposed method yields competitive fusion results when compared with existing state-of-the-art methods.