Image Colored-Pencil-Style Transformation Based on Generative Adversarial Network

Generative Adversarial Network (GAN) has an excellent performance in “ Image-to-Image Translation” field, this paper proposes an effective algorithm to solve the problem of image colored-pencil-style transformation based on Pix2Pix of GAN. The improved generator and discriminator are designed to be suitable for high resolution image generation. In the loss function, TV Loss is using for improving the smoothness of images and removing noise and overlapping shadows. In order to adapt to the practical application requirement of colored-pencil-style transformation, our data set comes from Flickr, the world's largest photo-sharing site, and includes high-resolution images of different types. The experimental results show that our proposed algorithm not only improves the resolution of images, but also makes the details clearer. Moreover, in all terms of objective indicators, our proposed algorithm is superior to Pix2Pix in different categories of data set. Therefore, through the comparison in visual subjectively and in indicators objectively, it is proved that the proposed algorithm has better performance than Pix2Pix.