Attention-based Pyramid Dilated Lattice Network for Blind Image Denoising

Mohammad Nikzad, Yongsheng Gao, Jun ZhouIJCAI 2021 · DOI

Introduction

Image denoising is a fundamental task in computer vision, aimed at recovering a clean image from a noisy observation. While deep learning has significantly advanced this field, two primary challenges persist: the need for "blind" denoising—where the noise level is unknown at inference time—and the trade-off between model efficiency and performance. High-performing models like Residual Dense Networks (RDN) often require millions of parameters, making them computationally expensive. Conversely, smaller models frequently lack the receptive field necessary to remove heavy noise without sacrificing fine details.

The Attention-based Pyramid Dilated Lattice Network (APDL-Net) addresses these issues through a novel architecture that captures multi-scale contextual information efficiently. By utilizing a unique triangular lattice structure with varying dilation rates, the model achieves state-of-the-art results with a significantly reduced parameter footprint.

The APDL-Net Architecture

The APDL-Net is designed around a sequence of Attention-based Pyramid Dilated Lattice (APDL) blocks. Each block integrates local feature extraction with global awareness to refine the denoising process.

The overall APDL-Net architecture, incorporating multiple APDL blocks, channel-wise attention, and global residual links.

The network employs Global Residual Learning (GR), allowing the model to focus specifically on learning the noise residual rather than the entire image structure. Within each APDL block, a Squeeze-and-Excitation mechanism (Channel-Wise Attention) is applied to prioritize important feature channels, helping the network distinguish between structural details and noise artifacts.

The Pyramid Dilated Lattice (PDL) Block

The core innovation of this work is the Pyramid Dilated Lattice (PDL) structure. Unlike standard sequential layers, the PDL arranges convolutional units in a triangular lattice where the vertical position ($h$) directly corresponds to the dilation rate ($d=h$).

Detailed view of the Pyramid Dilated Lattice (PDL) block, showing the triangular flow and varying dilation rates (h).

This hierarchical arrangement allows the network to process information at multiple scales simultaneously:

  • Expanding Receptive Fields: In the first half of the lattice, receptive fields expand to capture broad contextual information.
  • Refining Details: In the second half, the fields narrow back down to refine local details.
  • Two-Step Bottlenecks: To maintain parameter efficiency, each unit uses a $1 \times 1$ convolution for channel reduction followed by a $3 \times 3$ dilated convolution. This "deep but thin" approach provides a large receptive field without a massive memory footprint.

Experimental Results

The authors evaluated APDL-Net on standard benchmarks including BSD500, Set12, and Kodak24. Key findings include:

  • Superior Performance: On the CBSD68 dataset, APDL-Net consistently outperformed established models like DnCNN and FFDNet across noise levels ranging from $\sigma = 5$ to $50$.
  • Extreme Efficiency: The APDL-Net(6,64) configuration achieved performance comparable to the RDN while using 14.3x fewer parameters (1.53M vs. 22M).
  • Inference Speed: The model demonstrated faster inference times than the Residual Non-local Attention Network (RNAN) while maintaining higher restoration quality.
  • Versatility: Beyond Gaussian noise, the model proved effective at removing JPEG compression artifacts, showing superior structural preservation (SSIM) compared to specialized models like ARCNN.

Conclusion

APDL-Net demonstrates that a structured, hierarchical approach to dilation can provide the benefits of dense feature re-usage without the associated computational costs. While the model excels at handling Gaussian noise and JPEG artifacts, future research may explore its performance on non-Gaussian, spatially variant "real-world" camera noise and its application to other restoration tasks like super-resolution.

Summary updated

BibTeX
@inproceedings{nikzad2021attention,
  author = {Mohammad Nikzad and Yongsheng Gao and Jun Zhou},
  title = {{Attention-based Pyramid Dilated Lattice Network for Blind Image Denoising}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/129},
}
Figure