基于EBGAN的图像风格化技术

为了解决传统图像风格化算法生成图像的多样性较差的问题，本文提出一种基于EBGAN(Energy-Based Generative Adversarial Net)的网络模型，即在鉴别器中引入能量函数思想，设计Autoencoder使其能分别针对真假输入产生不同重构结果，计算输入图像重构前后的误差值，以此误差值作为能量概念用来鉴别输入图像。在Autoencoder的编码阶段，对于编码后的向量引入正交控制，控制同一batch中的两两向量最大正交化，推动生成器生成朝着不同方向发展的图像。使用该模型在Facades和Cityscapes数据集上进行实验，实验结果表明本文的网络模型能有效完成图像风格化过程，较传统图像风格化网络模型能生成更加多样化的图像。

An Image Style Conversion Technology Based on EBGAN

In order to solve the problem of poor diversity of the generated images in the traditional image style conversion algorithm, this paper proposes a network model based on EBGAN (Energy-Based Generative Adversarial Net). The idea of energy function is introduced into the discriminator, and the Autoencoder is designed to generate different reconstruction results for the true and false input respectively. The error value before and after the reconstruction of the input image is calculated, which is used as the energy concept to identify the input image. In the coding stage of Autoencoder, the orthogonal control is introduced in to the encoded vectors to control the maximum orthogonalization of two vectors in the same batch, so as to promote the generator net to generate images in different directions. Experiments on Facades and Cityscapes datasets show that the proposed network model can effectively achieve process of image stylization and generate more diversified images than the traditional network model.