Research on the Influence of Multi-Scene Feature Classification on Ink and Wash Style Transfer Effect of ChipGAN

Abstract

Ink painting represents a distinguished traditional art form in China. In recent years, numerous researchers have employed deep learning techniques to investigate ink wash style transfer. However, there has been a notable oversight regarding the impact of multi-scene feature classification on both the transfer effect and its evaluation. This paper aims to address this gap by employing multi-scene feature classification datasets to facilitate the transformation of images into ink wash style utilizing ChipGAN. In our study, we utilized a content dataset comprising photographic images of the Forbidden City and frames from documentary videos, characterized by varying structural complexities and weather conditions. Additionally, we compiled style datasets consisting of five distinct ink styles, collectively forming a comprehensive multi-scene feature classification dataset. These datasets will serve as the foundation for ChipGAN-based ink wash style transfer. We investigated the influence of the structural complexity and weather conditions of images prior to transfer on the transfer effect across different style datasets. Furthermore, we examined the impact of the three components of the Structural Similarity Index (SSIM) on the evaluation of ink wash style transfer. Through a combination of SSIM analysis and subjective visual quantification, our findings indicate that the structural complexity of images prior to transfer significantly affects the transfer effect across various style datasets. Moreover, within different style datasets, the three components of SSIM exert varying degrees of influence on the evaluation outcomes in the context of multi-scene feature classification. This research lays a robust foundation for the development of an evaluation system tailored to ink wash style transfer within the framework of multi-scene feature classification based on ChipGAN. Additionally, it offers an objective and efficient methodology for assessing the quality of ink wash images generated by ChipGAN in future applications. In the long term, this work may also provide a novel perspective for the intelligent evaluation of images across other artistic styles.

Keywords

• Ink and wash style transfer
• muti-scene feature classification
• SSIM indicators
• transfer effect