基于生成对抗网络的图像艺术风格迁移

Image stylization transfer using generative adversarial neural networks

提出了一种新颖的图像艺术风格化算法,利用结构相似性指数和最小二乘生成对抗网络,搭建图像艺术风格化模型.通过对模型生成器和判别器的对抗训练以及重建约束,该模型可以生成一幅逼真的风格化作品.根据在人脸肖像素描sketch-photo数据集和中国水墨画风格beihong-photo数据集实验表明,与目前流行的DualGAN算法、CycleGAN算法、Pix2Pix算法和GAN算法相比,本文提出的方法具有更好的风格化效果.

This paper proposes a novel neural style transfer algorithm. The Structural Similarity Index Measurement(SSIM) with the Least Squares Generate Adversarial Nets(LSGAN) is adopted to build a model for image repainting with neural styles. Through the adversarial training and reconstruction constraints, the model can generate a more realistic and reasonable stylized picture. Two groups of experiments are performed with CHUK sketch-photo dataset and a Chinese ink-painted style dataset(named beihong-photo which is constructed by the authors’ lab). Experimental results show that the presented method achieves a more acceptable state-of-art effect compared with those from the most popular algorithms, including DualGAN, Cycle GAN, Pix2 Pix and the classic GAN.