杨倩,,
陈小朋,
苑玉彬,
张泓国,
王霖
兰州交通大学电子与信息工程学院 兰州 730070
基金项目:国家自然科学基金(61861025)
详细信息
作者简介:沈瑜：女，1982年生，教授，硕士生导师，研究方向为深度学习、神经网络、图像处理
杨倩：女，1995年生，硕士，研究方向为神经网络、风格迁移
通讯作者:杨倩 13662175532@163.com
中图分类号:TN911.73; TP391计量
文章访问数:143
HTML全文浏览量:138
PDF下载量:26
被引次数:0
出版历程
收稿日期:2020-03-25
修回日期:2021-01-30
网络出版日期:2021-07-21
刊出日期:2021-08-10
Structural Refinement of Neural Style Transfer
Yu SHEN,Qian YANG,,
Xiaopeng CHEN,
Yubin YUAN,
Hongguo ZHANG,
Lin WANG
College of Electronics and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
Funds:The National Natural Science Foundation of China (61861025)
摘要
摘要:风格迁移过程中风格元素均匀分布在整个图像中会使风格化图像细节模糊，现有的迁移方法主要关注迁移风格的多样性，忽略了风格化图像的内容结构和细节信息。因此，该文提出结构细化的神经风格迁移方法，通过增加边缘检测网络对内容图像的轮廓边缘进行提取实现风格化图像内容结构的细化，凸显内容图像中的主要目标；通过对转换网络中的常规卷积层的较大卷积核进行替换，在具有相同的感受野的条件下，使网络模型参数更少，提升了迁移速度；通过对转换网络中的常规卷积层添加自适应归一化层，利用自适应归一化在特征通道中检测特定样式笔触产生较高的非线性同时保留内容图像的空间结构特性来细化生成图像的结构。该方法能够细化风格化图像的整体结构，使得风格化图像连贯性更好，解决了风格纹理均匀分布使得风格化图像细节模糊的问题，提高了图像风格迁移的质量。
关键词:图像处理/
深度学习/
神经网络/
风格迁移/
边缘检测/
归一化
Abstract:In the process of style transfer, stylized image details are blurred when style elements are evenly distributed in the whole image. Besides, the existing style transfer methods mainly focus on the diversity of transferred styles, ignoring the content structure and details of the stylized images. To this end, a neural style transfer method of structure refinement is proposed, which refines the content structure of stylized image by adding edge detection network to extract the contour edge of the content image to highlight the main objectives in the content image. By replacing the larger convolution kernel of the conventional convolution layer in the transfer network, the model parameters of the transfer network are reduced, and the transfer speed is improved, while ensuring that the original receptive field is unchanged. Through the adaptive normalization of the conventional convolution layer, the structure of the generated image is refined by using the adaptive normalization to detect certain style of stroke in the feature channel to produce high nonlinearity while preserving the spatial structure of the content image. The method can refine the overall structure of the stylized image, make the stylized image more coherent, that the stylized image details are blurred due to the uniform distribution of style texture, and improve the quality of image style transfer.
Key words:Image processing/
Deep learning/
Neural network/
Style transfer/
Edge detection/
Normalization
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=4018cad9-29a2-455f-a839-ef5285b4dbef
