Abstract
This paper presents a novel image denoising framework that integrates the Lifting Stationary Wavelet Transform (LSWT) with Gaussian Scale Mixture (GSM) modeling. The approach leverages the shift-invariance of LSWT to mitigate artifacts while employing GSM to capture the statistical dependencies of wavelet coefficients in local neighborhoods. Bayes Least Square (BLS) estimation is utilized to derive clean coefficients from noisy observations. Experimental results on standard benchmark images demonstrate that the proposed GSM-LSWT method consistently outperforms traditional wavelet-based techniques in both Peak Signal-to-Noise Ratio (PSNR) and visual quality.
Introduction
Image denoising is a fundamental task in digital image processing, essential for restoring image quality degraded by noise during acquisition or transmission. The primary objective is to suppress noise while preserving critical structural information such as edges and textures. While standard Wavelet Transforms (WT) are widely used, they often face challenges related to computational efficiency and shift-variance. The Lifting Scheme was introduced to enhance efficiency; however, the standard Lifting Wavelet Transform (LWT) lacks shift-invariance, leading to undesirable "pseudo-Gibbs" ringing artifacts around edges. This paper proposes a shift-invariant lifting-based approach combined with sophisticated statistical modeling to overcome these limitations.
Methodology
The proposed GSM-LSWT denoising pipeline consists of three main stages:
- Lifting Stationary Wavelet Transform (LSWT): To achieve shift-invariance, the traditional "split" step of the lifting scheme is omitted, and filters are modified using zero-padding. This ensures that the transform is robust to small translations in the input image.
- Gaussian Scale Mixture (GSM) Modeling: The statistical properties of wavelet coefficients are modeled using GSMs. A GSM represents a random vector as the product of a zero-mean Gaussian vector and an independent positive scalar (the scale). This model effectively captures the heavy-tailed marginal distributions and dependencies between neighboring coefficients.
- Bayes Least Square (BLS) Estimation: Noise reduction is performed by estimating the original coefficients using a BLS estimator. This process functions as a spatially adaptive local Wiener filter, adjusting its response based on the local image structure identified by the GSM model.
Finally, the denoised image is reconstructed using the inverse LSWT.
Experimental Results and Discussion
The performance of the GSM-LSWT method was evaluated using the standard "Lena," "Barbara," and "House" benchmark images. The study utilized the 'db1' (Haar) wavelet with four levels of decomposition.
Quantitative results are summarized in Table 1. Across all tested noise variances (σ = 20, 30, 40), the GSM-LSWT method consistently achieved higher PSNR values compared to Soft-thresholding (Soft-Wavelet), GSM in the traditional wavelet domain (GSM-Wavelet), and GSM in the Contourlet domain (GSM-Contourlet). For example, on the "Lena" image with σ = 20, the proposed method reached 32.92 dB, significantly outperforming the Soft-Wavelet baseline of 30.72 dB.
Qualitative analysis further highlights the advantages of the proposed approach. Figure 1 provides a visual comparison of the denoising results. The GSM-LSWT method preserves more fine-scale textures and produces fewer ringing artifacts compared to the other methods, particularly visible in the complex patterns of the "Barbara" image.
Conclusion
The combination of shift-invariant LSWT and GSM-based statistical modeling provides an effective framework for image denoising. By addressing the shift-variance of traditional lifting schemes and leveraging neighborhood context, the GSM-LSWT method achieves superior performance in noise suppression and feature preservation. Future research will explore the application of this framework to non-Gaussian noise models and further optimization of its computational complexity.
Summary updated