此应用程序的某些内容目前无法使用。
如果这种情况持续存在,请联系我们反馈与联系
1. (CN106651766) 一种基于深度卷积神经网络的图像风格迁移方法

专利局 : 中国
申请号: 201611252944.3 申请日: 30.12.2016
公布号: 106651766 公布日: 10.05.2017
公布类型: A
国际专利分类:
G06T 3/00
G PHYSICS
06
计算;推算;计数
T
一般的图像数据处理或产生
3
在图像平面内的图形图像转换,例如,从位像到位像地建立一个不同图像
申请人: SHENZHEN WETECCTV SCIENCE & TECHNOLOGY CO., LTD.
发明人: XIA CHUNQIU
优先权数据:
标题: (EN) Image style migration method based on deep convolutional neural network
(ZH) 一种基于深度卷积神经网络的图像风格迁移方法
摘要: front page image
(EN) The invention discloses an image text description method based on a visual attention model. The main content comprises the followings: image inputting, loss function training, stylizing, image enhancing and image thinning; and the processes are as follows: an input image is firstly adjusted as a content image (256*256) with a dual-linear down-sampling layer, and then stylized through a style subnet; and then a stylized result as the first output image is up-sampled as an image in the size of 512*512, and then the up-sampled image is enhanced through an enhancement subnet to obtain the second output image; the second output image is adjusted as the image in the size of 1024*1024, and finally, a thinning subnet deletes locally pixelated artifact and further thins the result to obtain a high-resolution result. By use of the image style migration method disclosed by the invention, the brushwork of the artwork can be simulated more closely; multiple models are combined into a network so as to process the image with bigger and bigger size shot by a modern digital camera; and the method can be used for training the combined model to realize the migration of multiple artistic styles.
(ZH) 本发明中提出的一种基于视觉注意模型的图像文字描述方法,其主要内容包括:图像输入、损失函数训练、风格化、图像增强、图像细化,其过程为,输入图像首先被调整为具有双线性下采样层的内容图像(256×256),通过风格子网风格化;接着,作为第一输出图像的风格化结果被上采样为512×512大小的图像,并且通过增强子网得到第二输出图像;然后,它被调整大小为1024×1024;最后,细化子网删除局部像素化伪影,并进一步细化结果,获得高分辨率结果。本发明可以更密切地模拟艺术品的笔触;将多个模型组合到一个网络中,能够处理现代数码相机拍摄得到的尺寸越来越大的图像;它还可以用于训练组合模型,实现多个艺术风格迁移。