Multiscale Structure Tensor for Improved Feature Extraction and Image Regularization.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Traditional regularization methods for image denoising struggle with high noise levels, often creating artifacts.
• Existing techniques using image gradients can fail to preserve salient features like edges and corners.

Purpose of the Study:

• To develop a robust image adaptive exponent-driven regularization method for filtering noisy images.
• To achieve salient feature preservation during the denoising process.

Main Methods:

• Utilized a spatially adaptive variable exponent function based on structure tensor eigenvalues.
• Developed a continuous switch mechanism for accurate identification of noisy edges and corners.
• Integrated the structure tensor-based exponent into a regularization model for image restoration.

Main Results:

• The proposed method effectively preserves image features and edges, even under high noise.
• GPU-based implementation achieves real-time edge map computation (45-60 frames/s).
• Outperformed L0 sparse gradient-based methods and other related filters in performance.

Conclusions:

• The multiscale structure tensor-based adaptive variable exponent method provides reliable edge maps.
• This approach offers robust and superior feature-preserving image denoising compared to existing methods.