Semantic-Related Image Style Transfer with Dual-Consistency Loss.
Neurocomputing(2020)
摘要
Recently, image style transfer has drawn great attention in computer vision area with its successful applications in digital entertainment and social media. Much of the research aimed at rendering an image with artistic style of single artwork. However in certain applications, the target is to render the image as it were painted by a certain artist. Even though collection-based image style transfer methods were proposed to address the single artwork problem, they ignored the relation between the subject and the style of an artwork, which is an important factor in mimicking an artist’s creation mechanism. Based on this consideration, we assume the semantic relationship between the content and style by constructing a subject-related photo-artwork dataset and seek to learn that relationship. We propose a dual-consistency loss to train an encoder decoder network with adversarial discriminator. The dual-consistency loss encourages the output to be semantically and stylistically consistent with subject-related content and style image pair. Thus the visual appearance of the stylized outputs would adapt to the subject of the content image, which increased the diversity of the generated results. Both quantitative and qualitative experimental results demonstrate the effectiveness of the proposed approach for retaining semantic and style consistency in style transfer process.
更多查看译文
关键词
Image style transfer,Generative adversarial network,Semantic and style consistency
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络