ConvCycleGAN - An Unsupervised Generative Adversarial Network for Chinese Pastoral Painting Style
Zhengyu Dong, Naixin He, Wang Li, Chaoqing Ma · 2024
The image style conversion work is to map the style of the target domain image while keeping the content of the natural landscape image in the source domain unchanged. This technology is widely used in fields such as painting, film, and artistic creation. At present, research in this field mostly focuses on learning the styles of Western painters, lacking research on traditional Chinese painting styles. Due to the limitations of various painting techniques and clear semantic boundaries in traditional Chinese painting, existing methods still cannot clearly express stroke boundaries. This article proposes a style conversion network based on CycleGAN named Conv-CycleGAN. Convnext blocks are added to the generator in combination with PONO-MS modules to extract finer boundary features, and a self-attention mechanism module is used in the discriminator to achieve high-quality image style conversion of Chinese pastoral landscape paintings. Through extensive experimental comparisons and evaluations, it has been proven that this method has the ability to optimize the traditional Chinese painting style, confirming its superiority.