FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining
Authors: Dong Li, Yidi Liu, Xueyang Fu, Jie Huang, Senyan Xu, Qi Zhu, Zheng-Jun Zha
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4. Experiment 4.1. Experimental Settings Datasets. For training, we employ the widely used Rain13k dataset (Chen et al., 2021). It contains 13,712 image pairs in the training set, and we evaluate the results on Rain100H (Yang et al., 2017), Rain100L (Yang et al., 2017), Test2800 (Fu et al., 2017b), and Test1200 (Zhang & Patel, 2018). Evaluation Metrics. Following previous work (Zamir et al., 2021; 2022), we adopt two commonly used quantitative metrics for evaluations: Peak Signal-to-Noise Ratio (PSNR) (Huynh-Thu & Ghanbari, 2008) and Structural Similarity Index (SSIM) (Wang et al., 2004). 4.2. Comparison with State-of-the-art Methods Table 1. Quantitative comparison (PSNR/SSIM) for Image Deraining on five benchmark datasets. Figure 5. Qualitative comparison on Rain100H (Yang et al., 2017). |
| Researcher Affiliation | Academia | 1University of Science and Technology of China, Hefei, China. Correspondence to: Zheng-Jun Zha <EMAIL>. |
| Pseudocode | No | The paper describes the methodology using textual explanations and mathematical equations, such as in Section 3 'Methodology' and its subsections '3.2. Scanning in Fourier Space' and '3.3. Fourier Mamba'. However, it does not include any distinct pseudocode blocks or algorithm listings. |
| Open Source Code | No | The paper does not contain an explicit statement or a link indicating that the source code for the described methodology is publicly available. It mentions implementation within PyTorch but no release. |
| Open Datasets | Yes | Datasets. For training, we employ the widely used Rain13k dataset (Chen et al., 2021). It contains 13,712 image pairs in the training set, and we evaluate the results on Rain100H (Yang et al., 2017), Rain100L (Yang et al., 2017), Test2800 (Fu et al., 2017b), and Test1200 (Zhang & Patel, 2018). To verify the generalization of the proposed method in real-world scenarios, we use the model trained on Rain13k to examine the real-world deraining capabilities. We evaluate the model trained on the synthetic dataset on the real-world dataset Inrernet-Data (Wang et al., 2019) without ground truth. To further explore the potential of the proposed method, we use the real-world dataset SPAData (Wang et al., 2019) to train Fourier Mamba. We use the LOL-V1 (Wei et al., 2018) and LOL-V2-synthetic (Wei et al., 2018) datasets to evaluate the performance of our method on low-light enhancement, and the Dense-Haze (Ancuti et al., 2019) and NH-HAZE (Ancuti et al., 2020) datasets are used to evaluate the performance of our method on real-world image dehazing. In addition, we also tested our method directly on Rain DS-Real (Quan et al., 2021). |
| Dataset Splits | Yes | For training, we employ the widely used Rain13k dataset (Chen et al., 2021). It contains 13,712 image pairs in the training set, and we evaluate the results on Rain100H (Yang et al., 2017), Rain100L (Yang et al., 2017), Test2800 (Fu et al., 2017b), and Test1200 (Zhang & Patel, 2018). |
| Hardware Specification | Yes | Our model is implemented within the Py Torch framework and executed on an NVIDIA A100 GPU. The comparison results of the model inference time using 512 × 512 images on NVIDIA RTX 4090 GPU are shown in Table 5. |
| Software Dependencies | No | The paper states, "Our model is implemented within the PyTorch framework" but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The number of blocks per layer has an impact on both the model s parameter count and its deraining performance. After balancing the weights, we configure the blocks per layer as [2, 3, 3, 4, 3, 3, 2], which allows us to achieve commendable performance with a reasonable number of parameters. We adopt the progressive training strategy. Specifically, we set the total number of iterations to 80,000 and image sizes to [160, 256, 320, 384], with the corresponding batch sizes of [8, 4, 2, 1]. We utilize the Adam optimizer with default parameters. The initial learning rate is established at 3 e 4, followed by a gradual decay to 1 e 6 using a cosine annealing schedule. |