多级注意力传播驱动的生成式图像修复方法

曹承瑞; 刘微容; 史长宏; 张浩琛

doi:10.16383/j.aas.c200485

多级注意力传播驱动的生成式图像修复方法

doi: 10.16383/j.aas.c200485

1.
兰州理工大学电气工程与信息工程学院兰州 730050

基金项目: 国家自然科学基金(61461028, 61861027)资助

详细信息

作者简介:
曹承瑞：兰州理工大学硕士研究生. 主要研究方向为深度学习和图像处理. E-mail: xiaocao1239@outlook.com

刘微容：兰州理工大学教授. 主要研究方向为机器视觉与人工智能, 复杂系统先进控制理论与应用. 本文通信作者. E-mail: liu_weirong@163.com

史长宏：兰州理工大学博士研究生. 主要研究方向为深度学习和图像处理. E-mail: changhong_shi@126.com

张浩琛：兰州理工大学电气工程与信息工程学院讲师. 主要研究方向为机器人传感与控制. E-mail: zhanghc@lut.edu.cn

计量
- 文章访问数: 1082
- HTML全文浏览量: 390
- PDF下载量: 214
- 被引次数: 0
出版历程
- 收稿日期: 2020-07-01
- 录用日期: 2020-12-14
- 网络出版日期: 2021-02-25
- 刊出日期: 2022-05-13

Generative Image Inpainting With Attention Propagation

1.
College of Electrical and Information Engineering, Lanzhou University of Technology, Lanzhou 730050

Funds: Supported by National Natural Science Foundation of China (61461028, 61861027)

More Information

Author Bio:
CAO Cheng-Rui　Master student at Lanzhou University of Technology. His research interest covers deep learning and image processing

LIU Wei-Rong　Professor at Lanzhou University of Technology. His research interest covers machine vision and artificial intelligence, advanced control theory and application. Corresponding author of this paper

SHI Chang-Hong　Ph. D. candidate at Lanzhou University of Technology. Her research interest covers deep learning and image processing

ZHANG Hao-Chen　Lecturer in the Department of Electrical and Information Engineering at Lanzhou University of Technology. His research interest covers sensing and control of robots

摘要

摘要: 现有图像修复方案普遍存在着结构错乱和细节纹理模糊的问题, 这主要是因为在图像破损区域的重建过程中, 修复网络难以充分利用非破损区域内的信息来准确地推断破损区域内容. 为此, 本文提出了一种由多级注意力传播驱动的图像修复网络. 该网络通过将全分辨率图像中提取的高级特征压缩为多尺度紧凑特征, 进而依据尺度大小顺序驱动紧凑特征进行多级注意力特征传播, 以期达到包括结构和细节在内的高级特征在网络中充分传播的目标. 为进一步实现细粒度图像修复重建, 本文还同时提出了一种复合粒度判别器, 以期实现对图像修复过程进行全局语义约束与非特定局部密集约束. 大量实验表明, 本文提出的方法可以产生更高质量的修复结果.
- 注意力传播 /
- 特征压缩 /
- 复合粒度判别器 /
- 图像修复
Abstract: There are disordered structures and blurred detail textures in most existing image inpainting methods, because the inpainting network cannot make full use of the information in the non-damaged regions to infer the contents of the damaged regions during reconstructing process. To address the issues, an image inpainting network driven by multi-scale attention propagation is proposed in this paper. Firstly, the high-level features extracted from the full-resolution images are compressed into multi-scale compact features, and then the compact features are driven to perform multi-scale attention feature propagation in order of scale size. As a result, the high-level features including structures and details are fully propagated in the network. In order to realize fine-grained image inpainting, a compound granularity discriminator is proposed to constrain image inpainting process with global semantic constraints and non-specific local dense constraints. A large number of experimental results show that the proposed method can restore higher quality inpainting results.
- Attention propagation /
- feature compression /
- compound granularity discriminator /
- image inpainting

HTML全文

图 1 当前图像修复方法所存在的结构和细节问题展示

Fig. 1 The structure and detail issues encountered in current image inpainting method

下载: 全尺寸图片幻灯片

图 2 常规自动编码器

Fig. 2 Conventional autoencoder

下载: 全尺寸图片幻灯片

图 3 多级注意力特征传播自动编码器

Fig. 3 Multi-scale attention propagation driven autoencoder

下载: 全尺寸图片幻灯片

图 4 多级注意力传播网络整体框架

Fig. 4 The framework of multistage attention propagation network

下载: 全尺寸图片幻灯片

图 5 注意力特征匹配与传播

Fig. 5 Flowchart of attention feature matching and propagation

下载: 全尺寸图片幻灯片

图 6 Places2数据集上的结果比较

Fig. 6 Comparisons on the test images from Places2 dataset

下载: 全尺寸图片幻灯片

图 7 Facade数据集上的结果比较

Fig. 7 Comparisons on the test image from Facade dataset

下载: 全尺寸图片幻灯片

图 8 CelebA-HQ数据集上的结果比较

Fig. 8 Comparisons on the test image from CelebA-HQ dataset

下载: 全尺寸图片幻灯片

图 9 有/无注意力传播时的图像修复结果

Fig. 9 Results with/without attention propagation

下载: 全尺寸图片幻灯片

图 10 有/无复合判别器时的图像修复结果

Fig. 10 Results with/without compound discriminator

下载: 全尺寸图片幻灯片

图 11 在Facade、CelebA-HQ和Places2数据集上的实例研究结果

Fig. 11 Case study on Facade, CelebA-HQ and Places2

下载: 全尺寸图片幻灯片

表 1 3个数据集的训练和测试分割

Table 1 Training and test splits on three datasets

数据集	训练	测试	总数
Facade	506	100	606
CelebA-HQ	28000	2000	30000
Places2	8026628	328500	8355128

下载: 导出CSV

表 2 CelebA-HQ、Facade和Places2数据集上的定量对比

Table 2 Quantitative comparisons on CelebA-HQ, Facade and Places2

数据集	掩码率	PSNR			SSIM			Mean L1 loss
数据集	掩码率	CA	MC	Ours	CA	MC	Ours	CA	MC	Ours
CelebA-HQ	10% ~ 20%	26.16	29.62	31.35	0.901	0.933	0.945	0.038	0.022	0.018
	20% ~ 30%	23.03	26.53	28.38	0.835	0.888	0.908	0.066	0.038	0.031
	30% ~ 40%	21.62	24.94	26.93	0.787	0.855	0.882	0.087	0.051	0.040
	40% ~ 50%	20.18	23.07	25.46	0.727	0.809	0.849	0.115	0.069	0.052
Facade	10% ~ 20%	25.93	27.05	28.28	0.897	0.912	0.926	0.039	0.032	0.028
	20% ~ 30%	25.30	24.49	25.36	0.870	0.857	0.871	0.064	0.052	0.047
	30% ~ 40%	22.00	23.21	24.53	0.780	0.815	0.841	0.084	0.068	0.059
	40% ~ 50%	20.84	21.92	23.32	0.729	0.770	0.803	0.106	0.086	0.074
Places2	10% ~ 20%	22.49	27.34	27.68	0.867	0.910	0.912	0.059	0.031	0.029
	20% ~ 30%	19.95	24.58	25.05	0.786	0.854	0.857	0.097	0.051	0.048
	30% ~ 40%	18.49	22.72	23.41	0.714	0.800	0.805	0.131	0.071	0.066
	40% ~ 50%	17.54	21.42	22.29	0.658	0.755	0.765	0.159	0.089	0.081

下载: 导出CSV

表 3 组件有效性研究

Table 3 Effectiveness study on each component

Att8	无	有	有	有	有	有
Att4	无	无	有	有	有	有
Att2	无	无	无	有	有	有
Att0	无	无	无	无	有	有
Single-D	有	有	有	有	有	无
Cg-D	无	无	无	无	无	有
Mean L1 loss	0.091	0.089	0.086	0.081	0.078	0.074

下载: 导出CSV

参考文献(29)

[1]	Wang Y, Tao X, Qi X, Shen X, Jia J. Image inpainting via generative multi-column convolutional neural networks. In: Proceedings of the 32nd Conference on Neural Information Processing Systems. Montreal, Canada: Curran Associates, Inc., 2018. 331−340
[2]	Yu J, Lin Z, Yang J, Shen X, Lu X, Huang T S. Generative image inpainting with contextual attention. In: Proceedings of the 2018 IEEE Conference on Computer Vision And Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 5505−5514
[3]	Barnes C, Shechtman E, Finkelstein A, Goldman D B. PatchMatch: A randomized correspondence algorithm for structural image editing. In: Proceedings of the ACM SIGGRAPH Conference. New Orleans, LA, USA: ACM, 2009. 1−11
[4]	Levin A, Zomet A, Peleg S, Weiss Y. Seamless image stitching in the gradient domain. In: Proceedings of the 8th European Conference on Computer Vision. Prague, Czech Republic: Springer, 2004. 377−389
[5]	Voronin V V, Sizyakin R A, Marchuk V I, Cen, Y, Galustov G G, Egiazarian K O. Video inpainting of complex scenes based on local statistical model. Electronic Imaging, 2016, 111(2): 681-690
[6]	Park E, Yang J, Yumer E, Berg A C. Transformation-grounded image generation network for novel 3D view synthesis, In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA: IEEE, 2017. 702−711
[7]	Simakov D, Caspi Y, Shechtman E, Irani M. Summarizing visual data using bidirectional similarity. In: Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition. Anchorage, AK, USA: IEEE, 2008. 1−8
[8]	Yeh R, Chen C, Lim T Y, Hasegawa J M, Do N M. Semantic image inpainting with perceptual and contextual losses. arXiv: 1607.07539v2, 2016.
[9]	Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde F D, Ozair S, et al. Generative adversarial nets. In: Proceedings of the 2014 Advances in Neural Information Processing Systems, arXiv: 1406.2661v1, 2014.
[10]	Mirza M, Osindero S. Conditional generative adversarial nets. arXiv: 14111784, 2014.
[11]	Iizuka Satoshi, Edgar. S-S, Hiroshi I. Globally and locally consistent image completion. ACM Transactions on Graphics, 2017, 36(4): 107:1-107:14
[12]	Song Y, Yang C, Lin Z, Liu X, Huang Q, Li H, et al. Contextual-based image inpainting: Infer, match, and translate. In: Proceedings of the 15th European Conference on Computer Vision. Munich, Germany: Springer, 2018. 3−18
[13]	Ballester C, Bertalmio M, Caselles V, Sapiro G, Verdera J, et al. Filling-in by joint interpolation of vector fields and gray levels. IEEE Transactions on Image Processing, 2001, 10(8): 1200-1211 doi: 10.1109/83.935036
[14]	Bertalmio M, Sapiro G, Caselles V, Ballester C, et al. Image inpainting. In: Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, New Orleans, LA, USA: SIGGRAPH, 2000. 417−424
[15]	Efros A A, Freeman W T. Image quilting for texture synthesis and transfer. In: Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, New York, NY, USA: SIGGRAPH, 2001. 341−346
[16]	朱为, 李国辉. 基于自动结构延伸的图像修补方法. 自动化学报, 2009, 35(8): 1041-1047 doi: 10.3724/SP.J.1004.2009.01041 ZHU Wei, LI Guo-Hui. Image Completion Based on Automatic Structure Propagation. ACTA AUTOMATICA SINICA, 2009, 35(8): 1041-1047 doi: 10.3724/SP.J.1004.2009.01041
[17]	王志明, 张丽. 局部结构自适应的图像扩散. 自动化学报, 2009, 35(3): 244-250 doi: 10.3724/SP.J.1004.2009.00244 WANG Zhi-Ming, ZHANG Li. Local-structure-adapted Image Diffusion. ACTA AUTOMATICA SINICA, 2009, 35(3): 244-250 doi: 10.3724/SP.J.1004.2009.00244
[18]	孟祥林, 王正志. 基于视觉掩蔽效应的图像扩散. 自动化学报, 2011, 37(1): 21-27 doi: 10.3724/SP.J.1004.2011.00021 MENG Xiang-Lin, WANG Zheng-Zhi. Image Diffusion Based on Visual Masking Effect. ACTA AUTOMATICA SINICA, 2011, 37(1): 21-27 doi: 10.3724/SP.J.1004.2011.00021
[19]	Pathak D, Krahenbuhl P, Donahue J, Darrell T, Efros A A. Context encoders: Feature learning by inpainting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Seattle, WA, USA: IEEE, 2016. 2536−2544
[20]	Liu G, Reda F A, Shih K J, Wang T, Tao A, Catanzaro b. Image inpainting for irregular holes using partial convolutions. In: Proceedings of the 15th European Conference on Computer Vision. Munich, Germany: Springer, 2018. 89−105
[21]	Zeng Y, Fu J, Chao H, Guo B. Learning pyramid-context encoder network for high-quality image inpainting. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA: IEEE, 2019. 1486−1494
[22]	Hao D, Neekhara P, Chao W, Guo Y. Unsupervised image-to-image translation with generative adversarial networks. arXiv: 1701.02676, 2017.
[23]	Gulrajani I, Ahmed F, Arjovsky M, Vincent D, Aaron C. Improved training of wasserstein gans. In: Proceedings of the 2017 Advances in Neural Information Processing Systems, Long Beach, CA, USA: Curran Associates, Inc., 2017. 5767−5777
[24]	Johnson J, Alahi A, Fei-Fei L. Perceptual losses for real-time style transfer and super-resolution. In: Proceedings of the 2016 European Conference on Computer Vision, Amsterdam, the Netherlands: Springer, 2016. 694−711
[25]	Gatys L A, Ecker A S, Bethge M. Image style transfer using convolutional neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Seattle, WA, USA: IEEE, 2016. 2414−2423
[26]	Gatys L, Ecker A S, Bethge M. Texture synthesis using convolutional neural networks. In: Proceedings of the 2015 Advances in Neural Information Processing Systems, Montreal, Quebec, Canada: Curran Associates, Inc., 2015. 262−270
[27]	Zhou B, Lapedriza A, Khosla A, Khosla A, Oliva A, Torralba A. Places: A 10 million Image Database for Scene Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(6): 1452-1464 doi: 10.1109/TPAMI.2017.2723009
[28]	Karras T, Aila T, Laine S, Lehtinen J. Progressive growing of GANs for improved quality, stability, and variation. arXiv: 1710.10196, 2017.
[29]	Tyleček R, Šára R. Spatial pattern templates for recognition of objects with regular structure. In: Proceedings of the 2013 German Conference on Pattern Recognition, Münster, Germany: Springer, 2013. 364−374