FFDNet: Toward a Fast and Flexible Solution for CNN-based Image Denoising **Abstract:** FFDNet is a novel fast and flexible denoising convolutional neural network designed to address the limitations of existing discriminative learning methods in image denoising. Unlike traditional methods that learn a specific model for each noise level, FFDNet uses a tunable noise level map as input, enabling it to handle a wide range of noise levels (0-75) with a single network. It also excels in removing spatially variant noise by specifying a non-uniform noise level map. FFDNet achieves faster speed than benchmark methods like BM3D on CPU without compromising denoising performance. Extensive experiments on synthetic and real noisy images demonstrate FFDNet's effectiveness and efficiency, making it suitable for practical denoising applications. **Keywords:** Image denoising, convolutional neural networks, Gaussian noise, spatially variant noise **Introduction:** Image denoising is crucial in low-level vision for improving visual quality and serving as a testbed for evaluating image prior models and optimization methods. Existing denoising methods, including model-based and discriminative learning approaches, have limitations in flexibility or efficiency. FFDNet overcomes these issues by incorporating a noise level map as input, allowing it to handle different noise levels and spatially variant noise with a single network. The network operates on downsampled sub-images to balance speed and performance, and orthogonal initialization is used to mitigate visual artifacts. **Related Work:** FFDNet builds on previous discriminative learning methods, which aim to learn image priors and inference parameters from training data. While these methods are efficient, they often require multiple models for different noise levels and cannot handle spatially variant noise. FFDNet addresses these limitations by explicitly modeling the relationship between noise level and denoising output. **Proposed Fast and Flexible Discriminative CNN Denoiser:** FFDNet is designed to be fast, flexible, and robust. It uses a noise level map to control the trade-off between noise reduction and detail preservation. The network operates on downsampled sub-images to improve efficiency and receptive field size. Orthogonal initialization is used to prevent visual artifacts when using higher noise levels. **Experiments:** FFDNet is evaluated on various datasets, including synthetic and real noisy images. Results show that FFDNet outperforms state-of-the-art methods in terms of denoising performance and computational efficiency. It also demonstrates superior performance in handling spatially variant noise and real-world noisy images. **Conclusion:** FFDNet is a promising solution for image denoising, offering fast inference, flexibility, and robustness. Its effectiveness and efficiency make it a valuable tool for practical denoising applications.FFDNet: Toward a Fast and Flexible Solution for CNN-based Image Denoising **Abstract:** FFDNet is a novel fast and flexible denoising convolutional neural network designed to address the limitations of existing discriminative learning methods in image denoising. Unlike traditional methods that learn a specific model for each noise level, FFDNet uses a tunable noise level map as input, enabling it to handle a wide range of noise levels (0-75) with a single network. It also excels in removing spatially variant noise by specifying a non-uniform noise level map. FFDNet achieves faster speed than benchmark methods like BM3D on CPU without compromising denoising performance. Extensive experiments on synthetic and real noisy images demonstrate FFDNet's effectiveness and efficiency, making it suitable for practical denoising applications. **Keywords:** Image denoising, convolutional neural networks, Gaussian noise, spatially variant noise **Introduction:** Image denoising is crucial in low-level vision for improving visual quality and serving as a testbed for evaluating image prior models and optimization methods. Existing denoising methods, including model-based and discriminative learning approaches, have limitations in flexibility or efficiency. FFDNet overcomes these issues by incorporating a noise level map as input, allowing it to handle different noise levels and spatially variant noise with a single network. The network operates on downsampled sub-images to balance speed and performance, and orthogonal initialization is used to mitigate visual artifacts. **Related Work:** FFDNet builds on previous discriminative learning methods, which aim to learn image priors and inference parameters from training data. While these methods are efficient, they often require multiple models for different noise levels and cannot handle spatially variant noise. FFDNet addresses these limitations by explicitly modeling the relationship between noise level and denoising output. **Proposed Fast and Flexible Discriminative CNN Denoiser:** FFDNet is designed to be fast, flexible, and robust. It uses a noise level map to control the trade-off between noise reduction and detail preservation. The network operates on downsampled sub-images to improve efficiency and receptive field size. Orthogonal initialization is used to prevent visual artifacts when using higher noise levels. **Experiments:** FFDNet is evaluated on various datasets, including synthetic and real noisy images. Results show that FFDNet outperforms state-of-the-art methods in terms of denoising performance and computational efficiency. It also demonstrates superior performance in handling spatially variant noise and real-world noisy images. **Conclusion:** FFDNet is a promising solution for image denoising, offering fast inference, flexibility, and robustness. Its effectiveness and efficiency make it a valuable tool for practical denoising applications.