研究目的
To develop a new space-variant regularisation term for variational image restoration that can handle different types of noise (additive white Gaussian noise, additive white Laplace noise, and salt and pepper noise) by automatically estimating per-pixel parameters from the observed image.
研究成果
The proposed space-variant regulariser, with automatically estimated parameters, achieves high-quality image restoration for various noise types and image characteristics, outperforming existing models in terms of ISNR and visual quality. It demonstrates flexibility in handling different gradient distributions and noise corruptions.
研究不足
The method assumes known blur and noise characteristics; parameter estimation may be affected by noise and blur, leading to potential inaccuracies. The ADMM convergence is not guaranteed for non-convex cases (pi < 1), and computational cost is high for large images.
1：Experimental Design and Method Selection:
The study uses variational methods with a new space-variant regulariser based on half-Generalised Gaussian distribution. The restoration is formulated as an optimization problem solved using the Alternating Direction Method of Multipliers (ADMM).
2：Sample Selection and Data Sources:
Synthetic grey level images (e.g., geometric, skyscraper, lungs, ecography, aneurism) are corrupted with known blur and noise (AWGN, AWLN, SPN). Images are sourced from repositories like https://medpix.nlm.nih.gov.
3：List of Experimental Equipment and Materials:
Computational setup for numerical simulations; no specific hardware mentioned, but involves software for image processing and optimization algorithms.
4：Experimental Procedures and Operational Workflow:
Parameters (p and α maps) are estimated from corrupted images using statistical inference. ADMM iterations are performed to minimize the cost functional, with stopping criteria based on convergence.
5：Data Analysis Methods:
Performance is evaluated using Improved Signal-to-Noise Ratio (ISNR) and Blurred Signal-to-Noise Ratio (BSNR). Visual inspections and numerical comparisons with baseline models (TV-L2, TVp-L2, TVsvp-L2 for AWGN; TV-L1, TVp-L1, TVsvp-L1 for impulsive noise) are conducted.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
