Explicit Filterbank Learning for Neural Image Style Transfer and Image Processing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing image style transfer methods often treat content and style as a black box.
• There is a need for a more explicit and understandable approach to neural style transfer.

Purpose of the Study:

• To propose a new, explicit, and decoupled perspective on image style transfer.
• To develop a method that allows for better understanding and control over style transfer.

Main Methods:

• Introduction of StyleBank, comprising multiple convolution filter banks, each representing a distinct style.
• Joint learning of StyleBank and an auto-encoder, ensuring the auto-encoder does not encode style information.
• Application of filter banks to intermediate features from the auto-encoder for style transfer.

Main Results:

• StyleBank enables explicit representation and control of styles.
• The method supports incremental learning for adding new styles and fusion of styles at image and region levels.
• Demonstrated comparable results to single-parameter methods in edge-aware image smoothing and denoising.

Conclusions:

• StyleBank offers a novel, interpretable framework for neural style transfer.
• The explicit decoupling provides new understanding and flexibility compared to traditional black-box methods.
• The general filter bank learning idea is applicable to various image processing tasks.