Iterative Residual Image Deconvolution

Siyao Li, Dongwei Ren, Furong Zhao, Zijian Hu, Junfeng Li, Qian Yin

Abstract: Image deblurring, a.k.a. image deconvolution, recovers a clear image from pixel superposition caused by blur degradation. Few deep convolutional neural networks (CNN) succeed in addressing this task. In this paper, we first demonstrate that the minimum-mean-square-error (MMSE) solution to image deblurring can be interestingly unfolded into a series of residual components. Based on this analysis, we propose a novel iterative residual deconvolution (IRD) algorithm. Further, IRD motivates us to take one step forward to design an explicable and effective CNN architecture for image deconvolution. Specifically, a sequence of residual CNN units are deployed, whose intermediate outputs are then concatenated and integrated, resulting in concatenated residual convolutional network (CRCNet). The experimental results demonstrate that proposed CRCNet not only achieves better quantitative metrics but also recovers more visually plausible texture details compared with state-of-the-art methods.

1. Introduction

Image deblurring, or deconvolution, is the process of recovering a sharp image from a blurry observation, typically assuming a known blur kernel. This task is inherently "ill-posed" because noise in the blurry image can be amplified during recovery, leading to artifacts like "ringing." While traditional methods use complex mathematical priors to maintain natural image statistics, they are often computationally slow. Conversely, standard Convolutional Neural Networks (CNNs) often struggle with deconvolution because they are not naturally designed for the mathematical inversion required. This paper introduces CRCNet (Concatenated Residual Convolutional Network), which bridges the gap between traditional signal processing and modern deep learning by unfolding the classical minimum-mean-square-error (MMSE) solution into a learnable residual architecture.

2. Methodology: From IRD to CRCNet

The authors' approach follows a clear logical progression from mathematical theory to architectural design. They first derive the Iterative Residual Deconvolution (IRD) algorithm by unfolding the MMSE solution into a power series of residual components. This shows that deconvolution can be viewed as an iterative process of extracting high-frequency details from the blurry input.

The CRCNet architecture consists of three main parts: channel expansion, an iterative residual part, and an integrative part.

This mathematical insight is translated into the CRCNet architecture, which consists of three stages:

  • Channel Expansion: A 1x1 convolution expands the input to a higher-dimensional feature space.
  • Iterative Residual Part: A sequence of Residual Units (RUs) mimics the iterative steps of the IRD algorithm. Unlike standard ResNets, these units are specifically designed to handle the "conv-deconv" structure found in the underlying math.
  • Integrative Part: Instead of a purely sequential flow, CRCNet concatenates the outputs of all previous residual units. An Integrative Unit (IU) then fuses these multi-scale residues to produce the final clear image.
Visual results of the IRD algorithm showing how image details are recovered as the number of iterations (N) increases.

3. Experimental Evaluation

The model was trained on a custom dataset of 860 sharp images from BSD500 and COCO, ensuring no pre-existing motion blur or bokeh effects. Evaluation was conducted using the standard Levin benchmark, a proposed synthetic set with random kernels, and real-world blurry photos captured with a GoPro Hero 6.

4. Key Results and Insights

CRCNet achieves state-of-the-art performance while remaining remarkably efficient. On the Levin benchmark, it reached a PSNR of 35.39 dB, outperforming other methods like FDN (35.08 dB) and IRCNN (34.68 dB). Notably, CRCNet uses only 0.4M parameters, compared to the 15M parameters required by the previous DCNN model.

Key insights from the study include:

  • Math-to-Architecture Mapping: The mathematical residues of an unfolded MMSE solution correspond directly to the residual learning blocks in the CNN, providing a theoretical foundation for the design.
  • The Power of Integration: Concatenating intermediate residues allows the network to leverage unique high-frequency information that might be lost in purely sequential architectures.
  • Visual Quality: Qualitatively, CRCNet produces images with sharper textures and fewer ringing artifacts, demonstrating robustness even on real-world data where blur kernels are estimated.

5. Conclusion

CRCNet offers an explicable and effective architecture for non-blind image deconvolution. By grounding the network design in the physics of the problem, the authors achieve superior results with significantly lower computational overhead. Future work may explore applying this "iterative-to-sequential" expansion to other low-level vision tasks such as denoising or super-resolution.

Summary updated

Abstract

Image deblurring, a.k.a. image deconvolution, recovers a clear image from pixel superposition caused by blur degradation. Few deep convolutional neural networks (CNN) succeed in addressing this task. In this paper, we first demonstrate that the minimum-mean-square-error (MMSE) solution to image deblurring can be interestingly unfolded into a series of residual components. Based on this analysis, we propose a novel iterative residual deconvolution (IRD) algorithm. Further, IRD motivates us to take one step forward to design an explicable and effective CNN architecture for image deconvolution. Specifically, a sequence of residual CNN units are deployed, whose intermediate outputs are then concatenated and integrated, resulting in concatenated residual convolutional network (CRCNet). The experimental results demonstrate that proposed CRCNet not only achieves better quantitative metrics but also recovers more visually plausible texture details compared with state-of-the-art methods.
BibTeX
@misc{li2018iterative,
  author = {Siyao Li and Dongwei Ren and Furong Zhao and Zijian Hu and Junfeng Li and Qian Yin},
  title = {{Iterative Residual Image Deconvolution}},
  year = {2018},
  eprint = {1804.06042},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1804.06042},
}
Figure