A model integrating attention mechanism and generative adversarial network for image style transfer

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image style transfer is crucial for tasks like image reconstruction and texture synthesis.
• Existing neural network models struggle with long-distance pixel dependencies inherent in style transfer.

Purpose of the Study:

• To develop a generation model for image style transfer that effectively addresses long-distance pixel dependencies.
• To enhance the model's ability to perceive and utilize relevant style and content features while suppressing irrelevant ones.

Main Methods:

• A generation model integrating a cycle-consistent network and an attention mechanism was constructed.
• The cycle-consistent mechanism facilitates mismatch conversion between input and output images.
• The attention mechanism improves perception of long-distance pixel dependencies and suppresses non-target style information.

Main Results:

• Experiments were conducted on the monet2photo dataset.
• The proposed model achieved a 45% misjudgment rate in Amazon Mechanical Turk (AMT) perceptual studies.
• This rate indicates a high level of perceptual quality and successful style transfer.

Conclusions:

• The cycle-consistent network model combined with an attention mechanism demonstrates significant advantages in image style transfer.
• The model effectively handles complex relationships between content and style features.
• This approach offers a promising solution for advanced computer vision applications requiring sophisticated image manipulation.