A practice-based exploration of post-processing adjustment method for creating variety in neural stylized images

Abstract

Neural style transfer (NST), a technique based on deep learning of convolution neural network (CNN) to create stylized pictures by stylizing ordinary pictures with the predetermined visual art style. In the past three years, NST has become a widely employed approach to produce various styles for the purpose of training in art education and industrial applications such as MoMA and Prisma. Whilst previously research is mainly focused on the production of abstract painting, the effect of NST is often visually impressive. However, the users argue that there are three issues should be carefully investigated during the generation of neural-stylized artwork which are the color scheme, the strength of style stroke, and the adjustment of contrast, which cannot meet the user needs. Based on the experiments of current NST-based methods, the author designed a post-processing software to validate the proposed method establish on image fusion, contrast enhancement, and blending technique which have been widely used in the processing of remote sensing images. The following are my research questions: 1) How to integrate BT, ICDDS and the blending technique into neural style transfer and provide more choices to users by generating adjusting parameters? 2) What is the value of selectivity offered by neural style transfer for artistic creation and art education? This thesis is a practice-based research that includes three phases: preliminary research, experiments, and evaluations. Preliminary research conducted with the iteration of projects, contextual review and reflective practice. Meanwhile, the author also conducted preliminary interview with design college students. Experiment involved testing of style transfer and automatic rendering machine. Evaluation including in-depth interview with experts to validate the proposed method for practical use in artistic creation and art education which show the value of trigger unconventional inspiration by using style transfer and automatic rendering machine in collect materials phase for artists and encourage students to develop personal unique style. In the light of this, since the method proposed in this research can provide multiple choices for the three issues in NST and no need to retrain, it will have certain application value in art education and industrial application.