A Generative Adversarial Network AMS-CycleGAN for Multi-Style Image Transformation

Abstract

The objective of image style transfer is to create an image that has the artistic features of a reference style image while also retaining the details of the original content image. Despite the promising outcomes of current approaches, they are still susceptible to generating image information distortion or noise texture problems due to the absence of an effective style representation. As a solution to the aforementioned issues, this paper proposes AMS-CycleGAN (Attention Moment Shortcut-Cycle Generative Adversarial Network), a CycleGAN-based method that achieves style transfer, resulting in artwork that closely resembles hand-painted masterpieces by artists. Initially, the framework makes use of the Positional Normalization-Moment Shortcut (PONO-MS) module, the purpose of which is to retain and transmit structural information in the generator. Additionally, the Multi-Scale-Structural Similarity Index (MS-SSIM) loss is added to strengthen the constraint on the brightness and colour contrast of images. Finally, an attention mechanism module is introduced in the discriminator to emphasize available features and suppress irrelevant features during the style transformation process. According to the experimental results obtained, our method demonstrates a higher level of consistency with human perception when compared to current state-of-the-art methods in image style transfer.

Keywords

• CycleGAN
• image style transfer
• multi-scale-structural similarity index
• attention mechanism