Map art style transfer with multi-stage framework
- PDF / 4,011,784 Bytes
- 15 Pages / 439.642 x 666.49 pts Page_size
- 48 Downloads / 204 Views
Map art style transfer with multi-stage framework Chiao-Yin Shih1 · Ya-Hsuan Chen1 · Tong-Yee Lee1 Received: 1 May 2020 / Revised: 13 August 2020 / Accepted: 28 August 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract We propose a multi-stage framework to create the stylized map art images. Existing techniques are successful in transferring style in photos. Yet, the noise in results and the harmonization in the generated art images still need to be investigated. We address these issues with a proposed algorithm that defines a good portrait for map art application in the initial round. A refinement strategy is then applied to produce the final map arts that meet the aforementioned expectations. Beside our plausible results, the objective evaluation presented in this paper shows that our proposed method can interactively achieve better and appealing map art results in the comparison with those of other works. In addition, our method can also create ocean or landscape stylized paintings using our map art collage. Keywords Map art images · NPR (Non-photorealistic rendering) · Stylization · Portraiture · Deep learning
1 Introduction From a style image and a content image, a new photo can be generated by transferring the reference style to the input content image. Producing such art images is not only an interesting research field but also a potential industry product. Modern artists generally use maps as background materials to create their multiple artworks. Robert Walden utilized the concept of ontology to create his ontology road maps (shown in Fig. 1a) and Ed Fairburn combines cartography and portraiture in his map art design. In particular, Ed Fairburn beautifully redraws and modifies intervenes with a range of original maps, especially creating gradual changes to contours, roads/streets, and other particular and natural patterns. These specially stylized drawing changes let his map art design to tease out the human feeling, giving a comfortable coexistence of photo and landscape and then to create appealing map art results such as Fig. 1b. However, such existing methods including conventional approaches and deep learning-based approaches remain challenges in the field of computer graphics, image processing, and multimedia. Tong-Yee Lee
[email protected] 1
National Cheng-Kung University, Tainan, Taiwan
Multimedia Tools and Applications
Fig. 1 Map art image examples
Our current study is motivated by the Neural Style algorithm by [5]. Given a style image and a content image, such a map art system [5] generates a new image that has the content and the style are transferred from the two input images. Unlike the introduced style transfer technique in [5], we propose a multi-stage framework. The proposed method aims to overcome the noise as well as to gain a better harmonization in the map art results. In the first stage, we use an reference style image and a content image to generate an initial map art. Thereafter, a proposed refinement strategy is employed to produce
Data Loading...
