体系化を目指した絵画風の画像生成

抄録

Recently, many researchers have implemented systems for nonphotorealistic rendering. It is because paintings are usually more effective than photographic images. But most of these focused on a particular painting style, for example, watercolor painting style, oil painting style and so on, or a particular painter's style, Van Gogh and so on. So, in this paper, we focus on a unified approach for paint-image generation. In concrete, we proposed an algorithm for converting the color of an input image according to the style of a reference painting.
The system is first given an input image and a reference painting. And it removes noise in the input image by median filtering. Next, it transfers all the pixel values of the CIE HLS color space in both the input and the reference and get the histograms. Then, we get the median of these histogram and shift histogram of the input image by the difference between the input and reference median. Next, we obtains spectrum of the wave of these histogram by using Fourier Transfer and mixes spectrums. Then, we make the new histogram by using Inverse Fourier Transfer of the composite spectrum. Finally, it converts the color of an input image according to the new histogram.
More, if users want to change an input image according to the style of a reference painting including edge line, we propose another algorithm for adding edge line to an input image. We convert an input image to grayscale and detect edge line by sobel filtering. Then, we compose the edge line and the new image which converted the color.
For this system, we can realize various style of painting. By analyzing and converting the color distribution, an input image can be changed to natural but painting-like color. Moreover, for example, this system can convert an image to the mixed style of painting, both watercolor and oil painting.