期刊
MULTIMEDIA TOOLS AND APPLICATIONS
卷 -, 期 -, 页码 -出版社
SPRINGER
DOI: 10.1007/s11042-023-15672-8
关键词
Generative Adversarial Network; Image Translation; Generative Model; Deep Learning
Recently, there has been significant progress in image-to-image translation in the literature. However, issues such as border distortion and color distortion continue to persist in existing methods. These methods often use multiple channels, which make it difficult to find the gradient in optimizer, resulting in unsatisfactory results. To address this problem, we propose the Three-Channel Generative Adversarial Network, which decomposes color images into three RGB color channels and utilizes single channel generators and dual discriminators for adversarial training. The algorithm also includes specific discriminators responsible for texture and structure discrimination, and a revised loss function to improve translation accuracy. Experimental results on various datasets demonstrate clear improvement over the pix2pix method in terms of quality and quantity.
Recently Image-to-image translation has achieve much progress in the literature. However, in exist method, border distortion, color distortion and others are the serious issues which continue to be resolved. Existing methods do not produce satisfactory results because the most exist methods are mainly used multi-channels which increase the difficulty of finding the gradient in optimizer. To address this problem, we proposed the Three-Channel Generative Adversarial Network. The algorithm decomposed color image into three color channels of RGB and utilized the single channel generators and dual discriminators in each color channel for adversarial training. The detailed discriminator adopted reversed PatchGAN which we proposed to be responsible for the image texture discrimination, and the structure discriminator adopted seven-layer convolutional structure to be responsible for the image structure discrimination. Then to improve the accuracy of translation, the loss function that associated to the network model has also been revised. In experiments, the ablation study were provided to prove the effectiveness of the algorithm on Cityscapes and Facades datasets. Our extensive experiments on a variety of datasets, including Style transfer, Image labeling transfer and End-to-end image dehazing, which consistently demonstrate clear improvement over the pix2pix method both qualitatively and quantitatively.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据