A Transformer-Based Model for Super-Resolution of Anime Image

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image Super-Resolution (ISR) enhances image quality but is underutilized in anime production.
• Current ISR methods predominantly use Convolutional Neural Networks (CNNs), with limited exploration of Transformer architectures.

Purpose of the Study:

• To develop a novel Anime Image Super-Resolution (AISR) method leveraging the Swin Transformer architecture.
• To improve the quality and resolution of anime images through advanced deep learning techniques.

Main Methods:

• Proposed an AISR method based on the Swin Transformer, incorporating shallow and deep feature extraction.
• Enhanced low-frequency information with a Gaussian filter and introduced variable window sizes, modifying Swin Transformer's patch merging.
• Utilized sub-pixel convolution for feature upsampling and channel expansion during image reconstruction.

Main Results:

• Trained and tested the AISR model on a custom high-quality anime dataset for various magnifications (2×, 4×, 8×).
• Achieved superior performance compared to conventional CNN-based and other transformer-based ISR methods.
• Demonstrated effectiveness through numerical metrics (PSNR, SSIM) and graphical comparisons.

Conclusions:

• The proposed Swin Transformer-based AISR method significantly outperforms existing approaches for anime image enhancement.
• The novel enhancements to the Swin Transformer architecture and the use of a dedicated anime dataset contribute to robust and high-quality results.