Structure-Texture Image Decomposition using Deep Variational Priors.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Traditional variational methods struggle to differentiate structure edges from high-contrast textures in image decomposition.
• Existing models often fail to distinguish fine-scale oscillations from true structural edges, leading to artifacts.

Purpose of the Study:

• To develop a new image decomposition model that learns deep variational priors for structure images without explicit training data.
• To improve the distinction between structure edges and texture oscillations using advanced deep learning techniques.

Main Methods:

• Introduced a novel model leveraging deep variational priors learned via convolutional neural networks (CNNs).
• Employed an alternating direction method of multipliers (ADMM) algorithm for integrating deep priors into iterative smoothing.
• Utilized CNNs as a replacement for traditional total variation priors to capture structure and texture characteristics.

Main Results:

• The learned CNN priors successfully differentiate high-amplitude details from structure edges, mitigating halo artifacts.
• The proposed formulation offers enhanced control over smoothing effects compared to previous data-driven methods.
• Demonstrated effectiveness across various applications including texture removal, detail manipulation, and tone-mapping.

Conclusions:

• Deep variational priors learned by CNNs offer a powerful new approach for structure-texture image decomposition.
• The method effectively addresses limitations of traditional variational formulations in distinguishing structural edges from textures.
• The model shows significant promise for diverse computational photography and image processing applications.