Non-Local Representation Based Mutual Affine-Transfer Network for Photorealistic Stylization

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Photorealistic stylization aims to create natural, camera-like images.
• Current methods struggle with global color consistency due to local stylization.

Purpose of the Study:

• To develop a method for photorealistic stylization with improved global color consistency.
• To address the limitations of existing local stylization techniques.

Main Methods:

• Proposed a non-local representation scheme constrained with a mutual affine-transfer network (NL-MAT).
• Utilized dictionary-based decomposition to decouple representations and color information.
• Incorporated context correspondence for natural style transfer.

Main Results:

• Achieved local style transfer with global color consistency.
• Preserved spatial structure and original color properties of the content image.
• Demonstrated superior photorealistic results compared to state-of-the-art methods.

Conclusions:

• The NL-MAT method offers a novel approach to photorealistic stylization.
• This is the first method to use non-local representations for this task without semantic matching.
• The technique successfully enhances realism by maintaining global color consistency.