Adaptive Dynamic Filtering Network for Image Denoising

Hao Shen, Zhong-Qiu Zhao, Wandi ZhangAAAI 2023 · DOI

TL;DR

This paper introduces ADFNet, an adaptive dynamic filtering network for image denoising that leverages dynamic convolution with a novel spatially enhanced kernel generation (SEKG) module. By effectively incorporating spatial context and multi-scale features, ADFNet achieves superior denoising performance with lower computational complexity compared to existing methods.

What Problem Does This Paper Address?

Traditional image denoising methods often rely on image priors, which can fail when noise distributions deviate from learned priors. While Convolutional Neural Networks (CNNs) have shown great promise, many deep CNN-based denoising methods suffer from high computational costs and can over-smooth images, losing crucial high-frequency details like edges and textures. Existing approaches that enlarge receptive fields or use dynamic convolutions often neglect spatial context or struggle with efficient multi-scale feature extraction. This paper aims to bridge this gap by developing a denoising network that effectively captures both high-frequency and multi-scale information while remaining computationally efficient.

Methods

The core of ADFNet is its novel Spatially Enhanced Kernel Generation (SEKG) module. This module enhances dynamic convolution by incorporating spatial context information into kernel generation with minimal computational overhead. It comprises two branches: a spatial context extractor and a channel information interaction branch, which are fused to guide the generation of dynamic kernels.

Building upon SEKG, the paper proposes two key blocks:

  • Dynamic Convolution Block (DCB): This block uses dynamic convolution (guided by SEKG) to enhance high-frequency information while preserving low-frequency information through skip connections.
  • Multi-Scale Dynamic Convolution Block (MDCB): This block expands the receptive field using dilated convolutions and extracts powerful multi-scale features by employing shared adaptive dynamic kernels. This sharing mechanism reduces computational complexity.

Additionally, a Multi-Dimension Feature Integration (MFI) mechanism is introduced to fuse multi-scale features by capturing cross-dimension inter-dependencies, leading to richer feature representations. The overall network architecture follows an encoder-decoder structure with residual connections. For efficiency, depth-wise separable convolutions are employed.

Datasets

  • Training:
    • DIV2K (for Gaussian image denoising)
    • Smartphone Image Denoising Data set (SIDD) (for real-world image denoising)
  • Testing:
    • BSD68, Kodak24, McMaster (for Gaussian image denoising)
    • Darmstadt Noise Dataset (DND), SIDD (for real-world image denoising)

No new benchmarks were introduced; existing datasets were used.

Key Insights and Ideas

  • Spatial Context is Crucial for Dynamic Kernels: The SEKG module demonstrates that explicitly incorporating spatial context into dynamic kernel generation significantly improves performance, especially for high-frequency details.
  • Dynamic Convolution for High-Frequency Details: Dynamic convolution, guided by SEKG, proves effective in restoring complex high-frequency information that static convolutions often miss.
  • Shared Kernels for Multi-Scale Efficiency: Using shared adaptive dynamic kernels in the MDCB is an efficient way to extract multi-scale features without a prohibitive increase in parameters or computation.
  • MFI for Richer Representations: The Multi-Dimension Feature Integration mechanism effectively fuses multi-scale features by considering inter-dependencies across different dimensions, leading to more robust representations.
  • Decoder-Focused Application is Optimal: Applying dynamic convolution modules primarily in the decoder (where noise is reduced) yields better results than applying them in the encoder, as noisy kernels can be detrimental.

Key Results

  • Quantitative Performance: ADFNet consistently achieves superior Peak Signal-to-Noise Ratio (PSNR) values on various Gaussian and real-world noisy datasets compared to state-of-the-art methods like RNAN, DeamNet, MPRNet, and DAGL. For example, on the Kodak24 dataset, ADFNet shows performance gains of up to 0.18 dB over DeamNet.
  • Computational Efficiency: ADFNet demonstrates significantly lower FLOPs and running times compared to many competing methods, offering a better trade-off between performance and computational cost. For instance, it achieves a better PSNR than MPRNet while running approximately 4 times faster.
  • Qualitative Results: Visual comparisons show that ADFNet preserves rich image details, avoiding the smoothness artifacts introduced by some other methods.
  • Backbone Integration: The proposed DConv and MDCB modules show performance improvements when integrated into existing network architectures like DnCNN and RDN, highlighting their general applicability.

Limitations and Future Work

  • Sensitivity to Noise: The paper notes that dynamic convolution is sensitive to noise, which is why the proposed modules are primarily deployed in the decoder.
  • Performance vs. Depth/Channels: While performance generally improves with more channels and depth, the gains decrease, suggesting that stacking too many modules might not always be beneficial.
  • Future Work: The authors suggest that the SEKG-guided modules can serve as general-purpose layers or feature extraction modules, implying potential for broader applications beyond denoising. The effectiveness of these modules in various backbone architectures also points to future research in integrating them into diverse network designs.

Summary updated

BibTeX
@inproceedings{shen2023adaptive,
  author = {Hao Shen and Zhong-Qiu Zhao and Wandi Zhang},
  title = {{Adaptive Dynamic Filtering Network for Image Denoising}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i2.25317},
}
Figure