Abstract:The discriminative model learning for image denoising has been recently attracting considerable attentions due to its favorable denoising performance. In this paper, we take one step forward by investigating the construction of feed-forward denoising convolutional neural networks (DnCNNs) to embrace the progress in very deep architecture, learning algorithm, and regularization method into image denoising. Specifically, residual learning and batch normalization are utilized to speed up the training process as well as boost the denoising performance. Different from the existing discriminative denoising models which usually train a specific model for additive white Gaussian noise at a certain noise level, our DnCNN model is able to handle Gaussian denoising with unknown noise level (i.e., blind Gaussian denoising). With the residual learning strategy, DnCNN implicitly removes the latent clean image in the hidden layers. This property motivates us to train a single DnCNN model to tackle with several general image denoising tasks, such as Gaussian denoising, single image super-resolution, and JPEG image deblocking. Our extensive experiments demonstrate that our DnCNN model can not only exhibit high effectiveness in several general image denoising tasks, but also be efficiently implemented by benefiting from GPU computing.
What problem does this paper attempt to address?
The paper attempts to address several key issues in image denoising:
1. **Improving Denoising Performance**: Traditional image denoising methods often require complex optimization processes, resulting in slow denoising speed and low computational efficiency. Additionally, these methods tend to rely on manually selected parameters, which limits their generalization ability. This paper proposes a method based on deep convolutional neural networks (DnCNN), which significantly improves denoising performance through residual learning and batch normalization techniques.
2. **Handling Unknown Noise Levels**: Most existing denoising models are trained for specific noise levels and cannot handle image denoising tasks with unknown noise levels (i.e., blind Gaussian denoising). The proposed DnCNN model can handle image denoising with unknown noise levels, exhibiting stronger robustness and generalization ability.
3. **Extending to Other Image Processing Tasks**: Besides Gaussian denoising, this paper also explores how to extend the DnCNN model to other image processing tasks, such as single image super-resolution (SISR) and JPEG image deblocking. By analyzing the connection between DnCNN and the trainable nonlinear reaction diffusion model (TNRD), the authors propose a unified framework that can handle multiple image denoising tasks using a single DnCNN model.
### Main Contributions:
1. **End-to-End Trainable Deep CNN**: This paper proposes an end-to-end trainable deep CNN model (DnCNN) for Gaussian denoising. Unlike existing methods that directly estimate the latent clean image, DnCNN adopts a residual learning strategy to predict the noise residual image.
2. **Combination of Residual Learning and Batch Normalization**: The study finds that residual learning and batch normalization can mutually promote each other, not only accelerating the training speed but also significantly enhancing the denoising performance.
3. **Handling General Image Denoising Tasks**: DnCNN can handle not only Gaussian denoising with specific noise levels but also blind Gaussian denoising, single image super-resolution, and JPEG image deblocking tasks, performing excellently in these tasks.
### Experimental Results:
Experiments show that DnCNN outperforms existing advanced methods such as BM3D, WNNM, and TNRD in Gaussian denoising tasks. For blind Gaussian denoising, DnCNN also performs excellently, even surpassing models trained for specific noise levels. Additionally, DnCNN achieves satisfactory results when extended to other image processing tasks.
In summary, this paper introduces an efficient deep CNN model by incorporating residual learning and batch normalization techniques, which not only excels in Gaussian denoising tasks but also handles various image processing tasks, showing broad application prospects.