Abstract: We introduce an integrodifferential extension of the edge-enhancing anisotropic diffusion model for image denoising. By accumulating weighted structural information on multiple scales, our model is the first to create anisotropy through multiscale integration. It follows the philosophy of combining the advantages of model-based and data-driven approaches within compact, insightful, and mathematically well-founded models with improved performance. We explore trained results of scale-adaptive weighting and contrast parameters to obtain an explicit modelling by smooth functions. This leads to a transparent model with only three parameters, without significantly decreasing its denoising performance. Experiments demonstrate that it outperforms its diffusion-based predecessors. We show that both multiscale information and anisotropy are crucial for its success.
Introduction
Image denoising has traditionally been split into two distinct methodologies. On one hand, Partial Differential Equation (PDE) models, such as Edge-Enhancing Anisotropic Diffusion (EED), offer mathematical transparency and compactness. However, these models typically operate on a single scale, limiting their ability to handle complex image structures. On the other hand, Deep Learning (DL) models provide state-of-the-art performance by leveraging multiscale information but often function as "black boxes" with millions of parameters and a lack of mathematical stability guarantees.
This paper introduces Integrodifferential Anisotropic Diffusion (IAD), a model that bridges this gap. By incorporating multiscale structural information into a mathematically rigorous PDE framework, the authors achieve high-performance denoising with extreme parameter efficiency.
Methodology: Integrodifferential Anisotropic Diffusion
The core innovation of IAD is the use of an integrodifferential extension to the standard EED model. While traditional EED relies on image gradients at a single smoothing level, IAD utilizes a multiscale structure tensor. This tensor integrates structural information across a continuous range of smoothing scales ($\sigma$).
This approach allows the model to create anisotropy—direction-dependent smoothing—through multiscale integration for the first time. By summing contributions from various scales, the model can robustly identify edge orientations even in the presence of significant noise, adapting its behavior to both fine details and broader image features.
Learning-Based Parameter Reduction
To optimize the model, the authors initially trained 96 parameters, representing scale-adaptive weights $\gamma(\sigma)$ and contrast parameters $\lambda(\sigma)$ across different noise levels. Analysis of these learned values revealed a critical insight: the parameters follow highly regular, smooth functional patterns across scales.
By modeling these dynamics with smooth functions, the authors reduced the 96-parameter set to just three global constants ($\alpha, \beta, \lambda_0$). This drastic reduction highlights the model's transparency; the "intelligence" of the system is captured in a few well-defined mathematical relationships rather than a massive weight matrix. Furthermore, the model remains stable in the Euclidean norm, ensuring predictable behavior.
Experimental Results
The IAD model was evaluated using 200 training and 100 test images from the BSDS500 dataset, corrupted with additive Gaussian noise.
Quantitative Performance
IAD consistently outperformed its predecessors. At a high noise level (standard deviation $s=50$), IAD achieved a Peak Signal-to-Noise Ratio (PSNR) of 24.61 dB, a notable improvement over the 24.24 dB achieved by traditional EED. Remarkably, the 3-parameter reduced model performed nearly as well as the 96-parameter version, with only a negligible 0.07 dB decrease in PSNR.
Qualitative Analysis
Visually, the IAD model excels at cleaning noisy backgrounds while maintaining sharp foreground boundaries. This is a significant improvement over single-scale models, which often struggle to distinguish between noise and true image structure in flat regions.
Conclusion
The Integrodifferential Anisotropic Diffusion model demonstrates that high-performance image processing does not necessitate "black box" complexity. By combining the strengths of model-based and data-driven approaches, the authors have created a transparent, stable, and highly efficient denoising tool. Future work may extend this integrodifferential logic to other domains such as image inpainting and segmentation, further exploring the potential of compact, mathematically well-founded learned models.
Summary updated