A discriminative approach for wavelet denoising.

Area of Science:

• Image processing and computer vision
• Signal processing
• Machine learning for image restoration

Background:

• Traditional wavelet denoising often relies on descriptive approaches requiring image or noise prior models.
• Existing methods like soft/hard thresholding have limitations, especially in over-complete representations.

Purpose of the Study:

• To propose a discriminative approach for wavelet denoising using learned mapping functions (MFs).
• To develop a method that bypasses the need for explicit image or noise prior modeling.
• To demonstrate the adaptability of the framework for various image restoration tasks.

Main Methods:

• A discriminative learning framework is employed to derive MFs directly from example image data.
• Least-squares fitting is utilized to train the mapping functions on an ensemble of images.
• The learned MFs are applied to wavelet transform coefficients for noise reduction.

Main Results:

• The developed MFs are fundamentally different from traditional thresholding techniques in over-complete settings.
• The proposed method achieves performance comparable to current state-of-the-art image denoising techniques.
• The framework demonstrates successful application to diverse restoration problems beyond simple denoising.

Conclusions:

• The discriminative wavelet denoising approach offers a powerful alternative to traditional methods.
• The learned mapping functions provide a flexible and effective tool for image restoration.
• This framework expands the applicability of shrinkage-based techniques to new challenges like deblurring and artifact removal.