Multiscale image blind denoising.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Traditional image denoising often relies on fixed noise models (e.g., Gaussian, Poissonian), which are insufficient for real-world images.
• Many images processed by end-users and scientists have unknown or imperfectly characterized noise due to complex imaging chains.
• Recent advancements allow noise model estimation from single images, considering signal and frequency dependencies.

Purpose of the Study:

• To develop a blind image denoising algorithm capable of handling complex, signal- and frequency-dependent noise.
• To adapt a multiscale denoising approach for unknown noise models.
• To evaluate the algorithm's performance on diverse real-world and simulated distorted images.

Main Methods:

• Proposed a multiscale denoising algorithm tailored for a broad range of noise models.
• Developed a blind denoising approach that estimates the noise model from the image itself.
• Validated the algorithm on real JPEG images, scanned old photographs, and simulated distorted images.

Main Results:

• Demonstrated effective denoising on real JPEG images with unknown noise characteristics.
• Successfully applied the algorithm to scans of old photographs, showcasing its utility for archival purposes.
• Verified the algorithm's consistency on simulated distorted images and compared its performance against prior state-of-the-art blind denoising methods.

Conclusions:

• The proposed multiscale blind denoising algorithm offers a robust solution for images with unknown noise models.
• This method broadens the applicability of image denoising techniques to real-world scenarios beyond idealized noise assumptions.
• The algorithm shows competitive performance compared to existing blind denoising techniques.