Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Style Transfer: From Stitching to Neural Networks

Xinhe Xu, Zhuoer Wang, Yihan Zhang, Yizhou Liu, Zhaoyue Wang, Zhihao Xu, Muhan Zhao, Huaiying Luo

TL;DR

The traditional style transfer method excels in creating artistic abstractions but can struggle with seamlessness, whereas the machine learning method preserves the integrity of foreground elements while enhancing the background, offering improved aesthetic quality and computational efficiency.

Abstract

This article compares two style transfer methods in image processing: the traditional method, which synthesizes new images by stitching together small patches from existing images, and a modern machine learning-based approach that uses a segmentation network to isolate foreground objects and apply style transfer solely to the background. The traditional method excels in creating artistic abstractions but can struggle with seamlessness, whereas the machine learning method preserves the integrity of foreground elements while enhancing the background, offering improved aesthetic quality and computational efficiency. Our study indicates that machine learning-based methods are more suited for real-world applications where detail preservation in foreground elements is essential.

Style Transfer: From Stitching to Neural Networks

TL;DR

The traditional style transfer method excels in creating artistic abstractions but can struggle with seamlessness, whereas the machine learning method preserves the integrity of foreground elements while enhancing the background, offering improved aesthetic quality and computational efficiency.

Abstract

This article compares two style transfer methods in image processing: the traditional method, which synthesizes new images by stitching together small patches from existing images, and a modern machine learning-based approach that uses a segmentation network to isolate foreground objects and apply style transfer solely to the background. The traditional method excels in creating artistic abstractions but can struggle with seamlessness, whereas the machine learning method preserves the integrity of foreground elements while enhancing the background, offering improved aesthetic quality and computational efficiency. Our study indicates that machine learning-based methods are more suited for real-world applications where detail preservation in foreground elements is essential.
Paper Structure (10 sections, 4 figures)

This paper contains 10 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Traditional Style Transfer Method
  4. Deep-learning-based Style Transfer
  5. Results
  6. Extensive Study on the Traditional Method with More Complex Style Images
  7. Comparative Study
  8. Conclusions
  9. Conclusions
  10. Future Work

Figures (4)

  • Figure 1: The Patch-based Traditional Style and Texture Method by Efros et al. textureTransfer. Such style transfer is primarily accomplished by overlaying randomly selected, style-matched samples onto the input image. When integrating various patches onto the input image, the minimum error boundary cut algorithm is employed to determine the optimal seam, depending on whether the overlaps are positioned horizontally, vertically, or diagonally.
  • Figure 2: Showcases the effect of one of the traditional style transfer method work on transferring image styles with simple styles. When zooming to check the details of the transferred image, we can clearly see the texture of the carpet is perfectly reflected on the generated image.
  • Figure 3: Illustrates the extensive analysis of the traditional style transfer method by Alexei A. Efros et al. textureTransfer on more complicated figures with more complicated styles. The patch size of the algorithm is set up as 5, 11, 16, and 20 respectively.
  • Figure 4: Comparison between the representative traditional style transfer method textureTransfer and the representative style transfer method ding2024style. (a) Input image. (b) Style image. (c) Generated image by the representative traditional style transfer method. (d) Generated by the representative deep-learning-based style transfer method.