Abstract
We propose a data-dependent denoising procedure to restore noisy images. Different from existing denoising algorithms which search for patches from either the noisy image or a generic database, the new algorithm finds patches from a database that contains only relevant patches. We formulate the denoising problem as an optimal filter design problem and make two contributions. First, we determine the basis function of the denoising filter by solving a group sparsity minimization problem. The optimization formulation generalizes existing denoising algorithms and offers systematic analysis of the performance. Improvement methods are proposed to enhance the patch search process. Second, we determine the spectral coefficients of the denoising filter by considering a localized Bayesian prior. The localized prior leverages the similarity of the targeted database, alleviates the intensive Bayesian computation, and links the new method to the classical linear minimum mean squared error estimation. We demonstrate applications of the proposed method in a variety of scenarios, including text images, multiview images and face images. Experimental results show the superiority of the new algorithm over existing methods.
Introduction
Image denoising typically relies on either internal or external data sources. Internal methods, such as BM3D, exploit self-similarity within the noisy image. However, these methods suffer from the "rare patch effect," where unique features cannot be effectively recovered because similar patches are unavailable within the same image. Conversely, external methods utilize massive, generic databases to learn image priors. While powerful, these are often computationally expensive and can be biased by irrelevant data.
This paper proposes a middle ground: Adaptive Image Denoising by Targeted Databases. By utilizing a small, highly relevant database (e.g., a specific font set for text or a set of face images for portraits), the framework achieves the benefits of external data without the dilution of generic datasets.
Methodology
The denoising process is framed as an optimal filter design problem consisting of two primary stages: basis selection and coefficient determination.
Basis Selection via Group Sparsity
To represent image patches efficiently, the authors seek a basis that promotes group sparsity. A key theoretical contribution is the proof that minimizing the $\ell_{1,2}$ norm for group sparsity is mathematically equivalent to performing Principal Component Analysis (PCA) on weighted patches. This provides a rigorous foundation for PCA-based denoising heuristics.
Localized Bayesian Prior
Once the basis is established, spectral coefficients are determined using a localized Bayesian prior. Unlike global models that attempt to characterize all possible natural images, this model assumes the clean patch originates from a Gaussian distribution centered specifically around patches in the targeted database. This approach avoids the "averaging out" of specific details common in generic models.
Refined Patch Selection
Standard k-Nearest Neighbors (kNN) often fails under high noise because the search is corrupted by noise patterns. To mitigate this, a "first-pass" refinement step is introduced, using an initial rough denoising (e.g., BM3D) to guide the search for more accurate external patches in the targeted database.
Experimental Results
The algorithm was evaluated across three scenarios: text images, multiview scenes (Middlebury dataset), and human faces (Gore database).
Performance in High Noise
The method demonstrates significant robustness to extreme noise. For text images at $\sigma = 100$, the proposed method achieved a PSNR of 22.20 dB, significantly outperforming BM3D's 16.68 dB. In multiview and face denoising tasks, the method consistently outperformed state-of-the-art competitors like EPLL and KSVD by 1–2 dB, while requiring fewer training samples.
Conclusion
The use of targeted databases effectively addresses the rare patch effect and provides a more powerful prior than generic global models. While the method's performance depends on the relevance of the external database, it offers a practical and efficient solution for specialized denoising tasks. Future work includes automating the selection of targeted databases and analyzing the minimum data requirements for guaranteed performance.
Summary updated