Texture Synthesis with Desired Visual Impressions Using Deep Correlation Feature

Atsushi Takemoto, Kensuke Tobitani, Yusuke Tani, Taishi Fujiwara, Yoichi Yamazaki, Noriko Nagata

Abstract: We propose a texture synthesis method with desired visual impressions, to realize product designs suitable for individual needs. First, we conducted an evaluation experiment and quantified visual impressions of texture images. After that, we modeled the relationship between the quantified visual impressions and the image feature extracted from the convolutional neural network. Based on the obtained model, we updated the features to have the desired visual impressions and generated texture image. Finally, we confirmed the effectiveness of our method by conducting a verification experiment.