结构细化的神经风格迁移

沈瑜; 杨倩; 陈小朋; 苑玉彬; 张泓国; 王霖

doi:10.11999/JEIT200211

结构细化的神经风格迁移

doi: 10.11999/JEIT200211

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: 1078
- HTML全文浏览量: 641
- PDF下载量: 74
- 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

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。
- 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.
- Image processing /
- Deep learning /
- Neural network /
- Style transfer /
- Edge detection /
- Normalization

HTML全文

图 1 风格迁移模型

下载: 全尺寸图片幻灯片

图 2 边缘提取过程示意图

下载: 全尺寸图片幻灯片

图 3 不同深度的边缘检测图

下载: 全尺寸图片幻灯片

图 4 转换网络结构

下载: 全尺寸图片幻灯片

图 5 不同卷积核风格迁移纹理对比

下载: 全尺寸图片幻灯片

图 6 损失函数对比图

下载: 全尺寸图片幻灯片

图 7 纹理比较

下载: 全尺寸图片幻灯片

图 8 本文算法迁移效果展示

下载: 全尺寸图片幻灯片

图 9 实验结果对比

下载: 全尺寸图片幻灯片

图 10 实验结果对比

下载: 全尺寸图片幻灯片

图 11 客观评价指标

下载: 全尺寸图片幻灯片

表 1 步长和感受野参数设置

Layer	Conv1_2	Conv2_2	Conv3_3	Conv3_4	Conv4_3	Conv4_4	Conv5_3	Conv5_4
步长	1	2	4	4	8	8	16	16
接受域	5	14	40	44	92	100	196	212

下载: 导出CSV

表 2 在BSDS500数据集上的客观评价指标

指标	ODS	OIS	AP
5层融合边缘检测图	0.760	0.784	0.800
6层融合边缘检测图	0.774	0.797	0.798
7层融合边缘检测图	0.777	0.788	0.814
8层融合边缘检测图	0.786	0.802	0.822

下载: 导出CSV

表 3 迁移网络改进前后参数量对比

对应卷积层参数量	特征图通道数	步长	卷积核尺寸，参数量	卷积核尺寸，参数量
Conv1	32	1	9×9, 159418368	2×5×5, 98406400
Conv2	64	2	3×3, 8856576	3×3, 8856576
Conv3	128	2	3×3, 4428288	3×3, 4428288
Resblock1-Resblock5	128		23454552	23454552
Nearest_Conv1	64	1/2	3×3, 57600	3×3, 57600
Nearest_Conv2	32	1/2	3×3, 73728	3×3, 73728
Conv4	3	1	9×9, 15552	2×5×5, 9600
总参数量			196.30×10⁶	135.29×10⁶

下载: 导出CSV

表 4 风格迁移算法运行时间比较(s)

方法		Gatys^[4]	Huang^[6]	Johnson^[10]	Liu^[18]	本文
图像尺寸	256×256	15.86	0.018	0.015	0.083	0.013
	512×512	54.85	0.065	0.05	0.141	0.038
	1024×1024	214.44	0.275	0.21	0.37	0.255

下载: 导出CSV

参考文献(18)

[1]	KYPRIANIDIS J E, COLLOMOSSE J, WANG Tinghuai, et al. State of the “Art”: a taxonomy of artistic stylization techniques for images and video[J]. IEEE Transactions on Visualization and Computer Graphics, 2013, 19(5): 866–885. doi: 10.1109/TVCG.2012.160
[2]	袁野, 贾克斌, 刘鹏宇. 基于深度卷积神经网络的多元医学信号多级上下文自编码器[J]. 电子与信息学报, 2020, 42(2): 371–378. doi: 10.11999/JEIT190135 YUAN Ye, JIA Kebin, and LIU Pengyu. Multi-context autoencoders for multivariate medical signals based on deep convolutional neural networks[J]. Journal of Electronics &Information Technology, 2020, 42(2): 371–378. doi: 10.11999/JEIT190135
[3]	GATYS L A, ECKER A S, and BETHGE M. A neural algorithm of artistic style[J]. arXiv preprint arXiv: 1508.06576, 2015.
[4]	GATYS L A, ECKER A S, and BETHGE M. Image style transfer using convolutional neural networks[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 2414–2423. doi: 10.1109/cvpr.2016.265.
[5]	LUAN Fujun, PARIS S, SHECHTMAN E, et al. Deep photo style transfer[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, USA, 2017: 6997–7005. doi: 10.1109/CVPR.2017.740.
[6]	HUANG Xun and BELONGIE S. Arbitrary style transfer in real-time with adaptive instance normalization[C]. 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017: 1510–1519. doi: 10.1109/iccv.2017.167.
[7]	LI Chuan and WAND M. Combining Markov random fields and convolutional neural networks for image synthesis[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 2479–2486. doi: 10.1109/CVPR.2016.272.
[8]	DUMOULIN V, SHLENS J, and KUDLUR M. A learned representation for artistic style[J]. arXiv preprint arXiv: 1610.07629, 2016.
[9]	CHEN Yang, LAI Yukun, and LIU Yongjing. CartoonGAN: generative adversarial networks for photo cartoonization[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 9465–9474. doi: 10.1109/CVPR.2018.00986.
[10]	JOHNSON J, ALEXANDRE A, and LI Feifei. Perceptual losses for real-time style transfer and super-resolution[C]. The 14th European Conference on Computer Vision, Amsterdam, Netherlands, 2016: 694–711. doi: 10.1007/978-3-319-46475-6_43.
[11]	王鑫, 李可, 宁晨, 等. 基于深度卷积神经网络和多核学习的遥感图像分类方法[J]. 电子与信息学报, 2019, 41(5): 1098–1105. doi: 10.11999/JEIT180628 WANG Xin, LI Ke, NING Chen, et al. Remote sensing image classification method based on deep convolution neural network and multi-kernel learning[J]. Journal of Electronics &Information Technology, 2019, 41(5): 1098–1105. doi: 10.11999/JEIT180628
[12]	CHEN Chunfu, FAN Quanfu, MALLINAR N, et al. Big-little net: an efficient multi-scale feature representation for visual and speech recognition[J]. arXiv preprint arXiv: 1807.03848, 2018.
[13]	WANG Xin, YU F, DOU Ziyi, et al. SkipNet: Learning dynamic routing in convolutional networks[C]. The 15th European Conference on Computer Vision, Munich, Germany, 2018: 420–436. doi: 10.1007/978-3-030-01261-8_25.
[14]	LIU Yun, CHENG Mingming, HU Xiaowei, et al. Richer convolutional features for edge detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(8): 1939–1946. doi: 10.1109/TPAMI.2018.2878849
[15]	XIE Saining and TU Zhuowen. Holistically-nested edge detection[J]. International Journal of Computer Vision, 2017, 125(1): 3–18. doi: 10.1007/s11263-017-1004-z
[16]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]. The 13th European Conference on Computer Vision, Zurich, Switzerland, 2014: 740–755. doi: 10.1007/978-3-319-10602-1_48.
[17]	SAIF M M and SVETLANA K. WikiArt emotions: an annotated dataset of emotions evoked by art[C]. The 11th International Conference on Language Resources and Evaluation, Miyazaki, Japan, 2018: 1225–1238.
[18]	LIU Xiaochang, CHENG Mingming, LAI Yukun, et al. Depth-aware neural style transfer[C]. The Symposium on Non-Photorealistic Animation and Rendering, California, USA, 2017: 4. doi: 10.1145/3092919.3092924.