Dark and Bright Channel Prior Embedded Network for Dynamic Scene Deblurring

Jianrui Cai, Wangmeng Zuo, Lei ZhangIEEE TIP 2020 · OpenReview

Abstract

Recent years have witnessed significant progress in convolutional neural networks (CNNs) for dynamic scene deblurring. While most CNN models are generally learned by reconstruction loss defined on training data, incorporating suitable image priors and regularization terms into the network architecture can boost deblurring performance. In this work, we propose a Dark and Bright Channel Priors embedded Network (ECPeNet) to plug channel priors into a neural network for effective dynamic scene deblurring. A novel trainable Extreme Channel Prior embedded Layer (ECPeL) is developed to aggregate both channel priors and blurry image representations, and a sparse regularization is introduced to regularize the model learning. Furthermore, we present an effective multi-scale network architecture, namely Image Full Scale Exploitation (IFSE), which works in both coarse-to-fine and fine-to-coarse manners for better exploiting information flow across scales. Experimental results on the GoPro and Köhler datasets show that our proposed ECPeNet performs favorably against state-of-the-art deep image deblurring methods in terms of both quantitative metrics and visual quality.

The Challenge of Dynamic Deblurring

Blind image deblurring—recovering a sharp image from a blurry one without knowing the specific "blur kernel"—is a classic ill-posed problem. Historically, researchers have relied on two distinct toolkits:

  1. Optimization-based methods: These rely on human-engineered "priors" (mathematical assumptions about clean images, such as the Dark Channel Prior). While flexible, they are computationally expensive and often struggle with complex, non-uniform blur.
  2. Deep Learning (CNNs): These learn direct mappings from blurry to sharp images. They are fast and handle complex motion well but often ignore the statistical "common sense" found in traditional priors, which can lead to blurry edges or unnatural artifacts.

ECPeNet bridges this gap by embedding "extreme channel priors"—the statistical properties of the darkest and brightest pixels—directly into a modern CNN architecture.

Methodology: ECPeNet and IFSE

The core innovation of this work lies in the integration of traditional image processing insights into a deep learning framework through two primary components: the Extreme Channel Prior embedded Layer (ECPeL) and the Image Full Scale Exploitation (IFSE) architecture.

The ECPeNet architecture features a multi-scale flow with specialized layers for prior embedding and feature mapping.

Extreme Channel Prior embedded Layer (ECPeL)

The ECPeL is a trainable layer that aggregates blurry image features with representations of the "extreme channels" (the darkest and brightest pixels in local patches). In a blurry image, dark pixels become less dark and bright pixels become less bright due to averaging. By enforcing sparsity on these channels during training, the network is forced to favor "sharp-looking" statistics, effectively narrowing the search space for the sharp solution.

Bidirectional Multi-scale Flow

Unlike traditional deblurring networks that use a simple coarse-to-fine approach, the IFSE architecture allows information to flow in both directions—coarse-to-fine and fine-to-coarse. This bidirectional flow maximizes the information shared across different image resolutions, ensuring that fine-scale details help inform coarse-scale estimations and vice versa.

Experimental Results

The authors evaluated ECPeNet on the GoPro and Köhler datasets, two standard benchmarks for realistic, spatially-variant motion blur.

Qualitative comparison between the blurry input, the ECPeNet result, and the ground truth, including visualizations of the extreme channel maps.

Quantitative and Qualitative Performance

  • State-of-the-Art Metrics: On the GoPro dataset, ECPeNet achieved a PSNR of 31.10 dB, outperforming previous state-of-the-art methods by a significant margin (0.84 dB over Tao et al.).
  • Visual Fidelity: Qualitatively, the results demonstrate much clearer text and sharper object boundaries. As shown in the comparison figures, ECPeNet effectively removes ringing artifacts and residual blur that often plague pure CNN-based approaches.
  • Efficiency: The model processes a 720p image in approximately 0.65 seconds on a single GPU, making it orders of magnitude faster than traditional optimization methods while maintaining superior quality.

Conclusion

ECPeNet demonstrates that traditional image priors are not obsolete in the age of deep learning; rather, they serve as powerful regularizers that can constrain and improve neural network performance. By combining the Extreme Channel Prior with a novel bidirectional multi-scale architecture, this work sets a new benchmark for dynamic scene deblurring, offering a robust solution for recovering high-quality images from complex motion blur.

Summary updated

Abstract

Recent years have witnessed the significant progress on convolutional neural networks (CNNs) in dynamic scene deblurring. While most of the CNN models are generally learned by the reconstruction loss defined on training data, incorporating suitable image priors as well as regularization terms into the network architecture could boost the deblurring performance. In this work, we propose a Dark and Bright Channel Priors embedded Network (DBCPeNet) to plug the channel priors into a neural network for effective dynamic scene deblurring. A novel trainable dark and bright channel priors embedded layer (DBCPeL) is developed to aggregate both channel priors and blurry image representations, and a sparse regularization is introduced to regularize the DBCPeNet model learning. Furthermore, we present an effective multi-scale network architecture, namely image full scale exploitation (IFSE), which works in both coarse-to-fine and fine-to-coarse manners for better exploiting information flow across scales. Experimental results on the GoPro and Köhler datasets show that our proposed DBCPeNet performs favorably against state-of-the-art deep image deblurring methods in terms of both quantitative metrics and visual quality.
BibTeX
@article{cai2020dark,
  author = {Jianrui Cai and Wangmeng Zuo and Lei Zhang},
  title = {{Dark and Bright Channel Prior Embedded Network for Dynamic Scene Deblurring}},
  journal = {IEEE Trans. Image Process.},
  year = {2020},
  url = {https://openreview.net/forum?id=fdgMVJRBt3},
}
Figure