Illustration image style transfer method design based on improved cyclic consistent adversarial network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Style transfer aims to render images in the artistic style of another.
• Existing methods struggle to balance style transformation with content preservation, especially for illustrations.

Purpose of the Study:

• To develop an efficient style transfer method for illustrations.
• To improve the expressiveness and realism of illustration images.
• To maintain artistic integrity during style transfer.

Main Methods:

• Combines attention mechanisms with cycle-consistent adversarial networks (CycleGAN).
• Introduces an improved attention module to adaptively highlight key visual elements.
• Utilizes capabilities of both image restoration and style transfer.

Main Results:

• Achieved high-quality style transfer with high accuracy (>95%) and detail retention.
• Demonstrated low loss function values (approaching 0.00) and minimal FID scores on benchmark datasets.
• Resulting images exhibit more vibrant colors and distinct line features.

Conclusions:

• The proposed method effectively transfers style to illustrations while preserving original features.
• The approach offers superior accuracy and detail preservation compared to existing methods.
• The technique shows practical applicability for enhancing illustration realism and expressiveness.