ConvUNeXt: A Lightweight Convolutional Neural Network for Watercolor Image Translation

Image-to-image transformation is the task of transforming an image from one domain to another. It includes the task of converting an image to an artistic style such as oil painting, tile art, and watercolor. These conversion techniques have rapidly advanced with the development of machine learning in recent years. This paper proposes a method for generating watercolor paintings using machine learning technologies. The proposed method uses conditional generative adversary networks (cGANs) to convert the style of input images from real to watercolor. Specifically, we propose a ConvUNeXt generator that generates high-quality watercolor images while preserving the color information of the original image. It employs an encoder-decoder structure with skip connections and introduces multiple ConvNeXt blocks to learn image features. Experimental results show that the proposed model significantly outperforms traditional rendering methods in the time consumption of image translation while achieving comparable translation performance.