文献类型：会议论文

英文题名：Research on Image-to-Image Translation with Capsule Network

作者：Ye, Jian[1];Chang, Qing[1];Jia, Xiaotian[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200030, Peoples R China

会议论文集：28th International Conference on Artificial Neural Networks (ICANN)

会议日期：SEP 17-19, 2019

会议地点：Tech Univ Munchen, Klinikum Rechts Isar, Munich, GERMANY

主办单位：Tech Univ Munchen, Klinikum Rechts Isar

语种：英文

外文关键词：Capsule Network; Conditional Generative Adversarial Network; Image-to-image translation

摘要：Deep learning technologies provide a unified translation framework for image-to-image translation. In particular, Convolution Neural Network (CNN) plays a decisive role because of its remarkable flexibility and performance. Recently, a new architecture called Capsule Network (CapsNet) was proposed to improve CNN. Capsule Network is able to preserve more input's information, especially location information by it's unique capsule structure and dynamic routing algorithm. In this paper, we propose Capsule conditional Generative Adversarial Network (CapscGAN) for performing image-to-image translation tasks. The proposed model utilizes CapsNet to encode image into one capsule called PixelCapsule and combines it with Markovian discriminator (PatchGAN) as discriminator. For suiting to translation tasks, we modify CapsNet's structure and activation function. Through a series of experiments, we analyze effect of CapsNet's activation function, dimension and application in discriminator. Multiple datasets' results demonstrate that our model has higher translation quality than convolutional image translation framework.