TL;DR
This paper introduces a novel "best basis" denoising algorithm using non-stationary wavelet packets, which adapt to the local frequency content of images, especially in textured areas. By optimizing the basis selection and incorporating a translation-invariant extension, the method significantly improves denoising performance compared to existing techniques, particularly for complex textures.
What Problem Does This Paper Address?
Traditional image denoising methods often rely on fixed representations like wavelets or wavelet packets. While effective for piecewise smooth signals, these fixed bases struggle to efficiently represent and denoise images with complex structures, such as oscillating textures and oriented edges. Different structures are best captured by different types of atoms (e.g., curvelets for edges, waveatoms for textures). This paper addresses the challenge of finding an adaptive representation that can optimally handle these diverse image characteristics, leading to better denoising results. Specifically, it aims to overcome the limitations of fixed bases by developing a method that can select the most suitable basis from a rich dictionary of non-stationary wavelet packets.
Methods
The core contribution is an algorithm for selecting a "best basis" within a dictionary of non-stationary (NS) wavelet packets. Unlike standard wavelet packets, NS wavelet packets allow filters to vary with both scale and position, enabling them to adapt to local frequency content.
- Quadtree Parameterization: The NS wavelet packet dictionary is structured using a quadtree. Each node in the tree represents a specific decomposition level and spatial location, and is associated with a filter.
- Best Basis Search: A novel best basis selection algorithm is proposed. It extends classical dynamic programming approaches to the NS setting by exploring the quadtree structure. This involves a bottom-up recursive selection process to identify the optimal filters at each level of the tree, minimizing a Lagrangian cost function that balances representation fidelity with complexity.
- Denoising: Once the optimal basis is found, the noisy image coefficients are hard-thresholded. The thresholding level is chosen to be asymptotically minimax optimal.
- Translation Invariant Denoising: To further improve results and reduce artifacts, a cycle spinning technique is employed. This involves applying the denoising process to translated versions of the image and averaging the results, making the denoising more robust to shifts.
Datasets
The paper evaluates its method on natural images, specifically mentioning the "Barbara" image and a "fingerprint texture." These are standard benchmark images used in image processing research for evaluating denoising algorithms. No new datasets were introduced.
Key Insights and Ideas
- Adaptivity is Key for Textures: The paper strongly demonstrates that adaptive representations, like NS wavelet packets, are crucial for effectively denoising images with complex textures, outperforming fixed bases.
- Non-Stationarity Matters: Allowing filters to vary with position (non-stationarity) is shown to be a significant advantage over standard wavelet packets for capturing local image characteristics.
- Cycle Spinning Boosts Performance: The translation invariant extension via cycle spinning provides a substantial improvement in denoising quality, particularly by reducing artifacts and enhancing performance on textured regions.
- Best Basis Selection as a Unified Framework: The "best basis" approach provides a principled way to adapt the signal representation to the image content, unifying denoising and representation.
Key Results
- Superior Denoising on Textures: The proposed NS wavelet packet denoising method (NSWP) achieved higher Peak Signal-to-Noise Ratios (PSNR) than standard wavelet packets and waveatoms on both the Barbara image and a fingerprint texture.
- On the Barbara image, NSWP achieved 25.4 dB, compared to WavePackets at 23.7 dB and WaveAtoms at 24.4 dB.
- On the fingerprint texture, NSWP achieved 24.5 dB, outperforming the original image (18.0 dB) and showing significant improvement.
- Translation Invariance Impact: The cycle spinning extension consistently improved denoising results, showing a 2 dB PSNR gain on the fingerprint texture.
- Computational Complexity: The complexity of the best basis algorithm is dominated by coefficient computation, which is shown to be O(n log(n)), where n is the number of pixels. The cycle spinning adds a multiplicative factor of K², where K is the number of translations.
Limitations and Future Work
- Computational Cost of Cycle Spinning: While effective, the cycle spinning technique increases the computational complexity by a factor of K². The paper uses K=4, leading to a 16x increase in the last step.
- Parameter Selection: The choice of the penalization parameter 'p' in the Lagrangian and the threshold 'T' for denoising are important. While the paper suggests methods for their selection (e.g., oracle manner for T), their optimal tuning can still be a challenge.
- Future Work: The authors suggest that this adaptive NS wavelet packet approach is competitive with state-of-the-art methods for denoising oscillating textures and implies potential for further improvements by exploring more sophisticated thresholding or basis selection criteria.
Summary updated