IFFMStyle: High-Quality Image Style Transfer Using Invalid Feature Filter Modules.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Image Processing

Background:

• Image style transfer aims to render images in various artistic styles.
• Existing methods often struggle with preserving content details and image structures when transferring styles from an artist's collection.
• Noise and texture elements can disrupt original image integrity in current approaches.

Purpose of the Study:

• To develop a high-quality image style transfer framework capable of capturing an artist's precise style from a collection of paintings.
• To address limitations of existing methods regarding content detail preservation and structure deformation.
• To enable style transfer without requiring paired content and style images.

Main Methods:

• Proposed IFFMStyle framework utilizing an encoder-decoder architecture.
• Introduced invalid feature filtering modules (IFFM) to remove content-independent features.
• Implemented content-consistency constraint and style perception consistency loss for enhanced content preservation and semantic distinction.

Main Results:

• IFFMStyle significantly improves the quality of stylized images.
• The method successfully transfers artistic styles based on semantic information of the content image.
• Experimental results indicate user preference for the proposed method over advanced techniques.

Conclusions:

• IFFMStyle offers a robust solution for high-quality image style transfer from an artist's collection.
• The framework effectively filters irrelevant features, preserving original image content and structure.
• This approach advances the field by enabling style transfer without paired data and achieving superior results.