Category-Specific Object Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Existing image denoising algorithms often rely on generic databases or self-similarity within noisy images.
• These methods can struggle with preserving fine details and object integrity.

Purpose of the Study:

• To develop a novel image denoising algorithm that utilizes external, category-specific image databases.
• To improve the performance of noisy image restoration by incorporating class-specific information.

Main Methods:

• The algorithm selects clean images from a category-specific database that are similar to the noisy input.
• It assembles 'support patches' from these clean images, corresponding to the noisy patch's spatial locality and object part.
• A content-adaptive distribution model is used, with parameters derived from support patches, and denoising is formulated as a transform-domain optimization problem.

Main Results:

• The objective function includes a Gaussian fidelity term with category-specific information and a low-rank term promoting patch similarity.
• The denoising process iteratively refines support patch selection and objective function optimization.
• Experiments across five object categories demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• Incorporating category-specific information significantly benefits image noise removal.
• The proposed method achieves superior denoising performance by leveraging external, class-relevant image data.