**FreqMamba: Viewing Mamba from a Frequency Perspective for Image Deraining** **Authors:** - Zhen Zou - Hu Yu - Feng Zhao **Abstract:** This paper introduces FreqMamba, an innovative method for image deraining that integrates frequency analysis with the Mamba model. The core of FreqMamba lies in extending Mamba with frequency analysis from two perspectives: frequency bands for exploiting frequency correlation and Fourier transform for global degradation modeling. The method introduces a triple interaction structure, including spatial Mamba, frequency band Mamba, and Fourier global modeling. Frequency band Mamba decomposes images into sub-bands of different frequencies, allowing 2D scanning from the frequency dimension. The method leverages Mamba's data-dependent properties to use rainy images at different scales for efficient training. Extensive experiments show that FreqMamba outperforms state-of-the-art methods both visually and quantitatively. **Keywords:** Image deraining, Frequency analysis, State space model **Introduction:** Image deraining is crucial for improving the quality of images corrupted by rain streaks, which often lose vital frequency information. Traditional methods rely on prior knowledge and physical models, while deep learning approaches have shown significant advancements. Mamba, a state space model, excels in modeling long-range sequence relationships with linear complexity. However, it lacks global degradation perception, which is crucial for image deraining. FreqMamba integrates Fourier transform and wavelet packet transform to enhance Mamba's global perception capabilities. The method uses a multi-scale U-Net architecture with three-branch Frequency-SSM blocks to address global degradation and refine local content. **Method:** FreqMamba's architecture includes a Fourier Modeling Branch, a Spatial Branch, and a Frequency Band Branch. The Fourier Modeling Branch processes the frequency domain representation of features, capturing global recovery. The Spatial Branch extracts details and correlations within the image. The Frequency Band Branch decomposes the input into sub-bands and processes them in the frequency dimension. The method also introduces a data-dependent degradation prior attention map, which enhances the model's ability to handle varying degrees of degradation across different regions of the image. **Experiments:** FreqMamba is evaluated on various datasets, including Rain100H, Rain100L, Test1200, and Test2800. Quantitative and qualitative comparisons with state-of-the-art methods show that FreqMamba achieves superior performance in both global and local recovery. Ablation studies further validate the effectiveness of each component of the model. **Conclusion:** FreqMamba integrates spatial domain sequence modeling and frequency domain global modeling to address image deraining effectively. It demonstrates robustness and versatility in various image restoration tasks, making it a promising foundation for future research in image restoration.**FreqMamba: Viewing Mamba from a Frequency Perspective for Image Deraining** **Authors:** - Zhen Zou - Hu Yu - Feng Zhao **Abstract:** This paper introduces FreqMamba, an innovative method for image deraining that integrates frequency analysis with the Mamba model. The core of FreqMamba lies in extending Mamba with frequency analysis from two perspectives: frequency bands for exploiting frequency correlation and Fourier transform for global degradation modeling. The method introduces a triple interaction structure, including spatial Mamba, frequency band Mamba, and Fourier global modeling. Frequency band Mamba decomposes images into sub-bands of different frequencies, allowing 2D scanning from the frequency dimension. The method leverages Mamba's data-dependent properties to use rainy images at different scales for efficient training. Extensive experiments show that FreqMamba outperforms state-of-the-art methods both visually and quantitatively. **Keywords:** Image deraining, Frequency analysis, State space model **Introduction:** Image deraining is crucial for improving the quality of images corrupted by rain streaks, which often lose vital frequency information. Traditional methods rely on prior knowledge and physical models, while deep learning approaches have shown significant advancements. Mamba, a state space model, excels in modeling long-range sequence relationships with linear complexity. However, it lacks global degradation perception, which is crucial for image deraining. FreqMamba integrates Fourier transform and wavelet packet transform to enhance Mamba's global perception capabilities. The method uses a multi-scale U-Net architecture with three-branch Frequency-SSM blocks to address global degradation and refine local content. **Method:** FreqMamba's architecture includes a Fourier Modeling Branch, a Spatial Branch, and a Frequency Band Branch. The Fourier Modeling Branch processes the frequency domain representation of features, capturing global recovery. The Spatial Branch extracts details and correlations within the image. The Frequency Band Branch decomposes the input into sub-bands and processes them in the frequency dimension. The method also introduces a data-dependent degradation prior attention map, which enhances the model's ability to handle varying degrees of degradation across different regions of the image. **Experiments:** FreqMamba is evaluated on various datasets, including Rain100H, Rain100L, Test1200, and Test2800. Quantitative and qualitative comparisons with state-of-the-art methods show that FreqMamba achieves superior performance in both global and local recovery. Ablation studies further validate the effectiveness of each component of the model. **Conclusion:** FreqMamba integrates spatial domain sequence modeling and frequency domain global modeling to address image deraining effectively. It demonstrates robustness and versatility in various image restoration tasks, making it a promising foundation for future research in image restoration.