Style Transfer Via Texture Synthesis.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Style transfer merges image styles with content, creating artistic mixtures.
• Convolutional Neural Networks (CNNs) have recently shown impressive results in style transfer.
• Texture synthesis algorithms offer an alternative but often yield less impressive results compared to CNNs.

Purpose of the Study:

• To propose a novel style transfer algorithm extending existing texture synthesis methods.
• To achieve stylized image quality comparable to CNN-based approaches.
• To improve consistency in preserving content while enhancing stylistic elements.

Main Methods:

• Modification of the Kwatra et al. (2005) texture synthesis algorithm.
• Development of techniques for maintaining content integrity in specific regions.
• Implementation of methods for generating rich stylistic details in other areas.

Main Results:

• The proposed algorithm produces visually pleasing and diverse stylized images.
• Results are competitive with state-of-the-art CNN style transfer methods.
• The algorithm demonstrates flexibility in processing various content and style image pairs.

Conclusions:

• The novel algorithm offers a competitive alternative to CNN-based style transfer.
• It provides a fast and flexible approach for artistic image synthesis.
• The method successfully balances content preservation with stylistic enhancement.