基于生成对抗网络的非织造布二维图像重建

摘要/Abstract

摘要： 非织造布纤维结构的准确表征是其性能分析的重要基础。为了解决基于近似模拟的表征结果中纤维形态、结构与真实样本不一致的问题，提出了一种基于生成对抗网络的非织造布二维图像重建方法。使用全自动光学显微镜对非织造布图像进行抓取，并在此基础上构建纤维生成对抗网络（Fiber generation adversarial network，FGAN）对图像样本进行建模。针对高分辨率图像重建时存在的失真问题，采用多尺度训练策略，同时引入权重多样性损失。采用图像质量评估指标FID作为实验评价指标，分别与DCGAN、WGAN-GP、BEGAN、PROGAN等生成模型进行对比实验。结果表明：FGAN重建的非织造布图像质量更高；消融实验证明，多尺度训练策略与权重多样性损失函数中FID数值分别降低24.52%、20.31%。FGAN模型的提出，使非织造布结构分析摆脱对近似模拟方法的依赖，提供了准确的纤维分布信息，对非织造布的质量评估、性能优化等应用具有重要意义。

关键词: 非织造布, 孔隙, 生成对抗网络, 多样性损失, 图像重建

Abstract: Nonwovens are used in a wide range of applications, including medical and health care, tourism, construction and waterproofing, and agriculture, for a variety of filtration materials, gas or particle adsorption materials, sound insulation, and more. The functions of nonwovens are closely related to the structure of their fiber collections. Based on the premise that nonwovens are randomly arranged in the layers of the fiber network, researchers usually use parametric simulation, fractal theory or image generation algorithms to simulate the fibers of nonwovens. However, the fiber structure generated in this way has the following problems: (1) The fiber morphology based on parametric simulation is linear and does not agree with the actual fiber curl state. (2) The fractal theory-based simulations cannot reconstruct the real nonwoven fabric structure concretely. (3) The single fiber morphology reconstructed based on the generation algorithm is different from the real state of the fiber morphology.
In order to reconstruct a realistic image of the nonwoven fiber structure, this paper constructs a generative adversarial network (FGAN) with a multiscale training strategy on the GAN base framework. In order to improve the stability of model training, a multi-scale training strategy is used to train the model, and PixelwiseNormalization layer and standardization layer are also introduced in the construction of the model. The reconstructed fiber structure is randomly arranged, which is consistent with the real nonwoven fiber arrangement. To increase the diversity of fiber structures in the generated images, a weight diversity loss WMI Loss is proposed. the model is trained for a long time, and the final generator is stable to generate nonwoven images that are consistent with the real images. Compared with other generative models, FGAN has better stability and the reconstructed nonwoven images have higher quality and more diverse fiber structures. Also, to verify the effectiveness of multiple-degree training strategy and weight diversity loss, ablation experiments are performed on the model. The experimental results show that the evaluation index FID of the model-generated images is reduced by 24.52% under the effect of the multiscale training strategy and by 20.31% under the effect of the weight diversity loss. Finally, in order to verify the structural consistency between the generated images and the real images, the average porosity of the real images is calculated to be 30.87% and the average porosity of the generated images is 30.02%, which are very close to each other. And the pore number distribution curves of the generated image and the real image have a high overlap. From the above analysis, it can be verified that the porosity and pore distribution of the nonwoven fabric images generated by the model are consistent with the real images.
FGAN can reconstruct high quality 2D images of nonwovens, from which detailed information on fiber distribution and morphology can be obtained. This information can be used for nonwoven performance analysis, optimization of production processes, etc. The method is able to reconstruct a diverse range of thin nonwoven structures, while the analysis of thicker nonwoven structures is yet to be investigated.

Key words: Nonwoven fabric, pore, generate adversarial networks, diversity loss, image generation

中图分类号:

TS 941.79

王志禄, 侯珏, 杨阳, 刘正. 基于生成对抗网络的非织造布二维图像重建[J]. 现代纺织技术, 2024, 32(1): 54-63.

WANG Zhilu , HOU Jue, YANG Yang, LIU Zheng. 2D Image reconstruction of nonwoven fabrics based on generative adversarial networks[J]. Advanced Textile Technology, 2024, 32(1): 54-63.

参考文献

[1]HAJIMOHAMMADI M, SOLTANI P, SEMNANI D, et al. Nonwoven fabric coated with core-shell and hollow nanofiber membranes for efficient sound absorption in buildings[J]. Building and Environment, 2022, 213: 108887.
[2]金关秀,祝成炎.孔隙形状对熔喷非织造布过滤品质的影响[J].上海纺织科技,2018,46(11):15-18.
JIN Guanxiu, ZHU Chengyan. Effect of pore shape on the filter quality of melt-blown nonwoven[J].Shanghai Textile Science & Technology, 2018,46(11):15-18.
[3]田伟,雷新,从明芳,等.纺粘非织造布制备工艺与性能的关系[J].纺织学报,2015,36(11):68-71.
TIAN Wei,LEI Xin,CONG Mingfang,et al. Relationship between manufacture process and performance for spunbond nonwoven[J].Journal of Textile Research,2015,36(11):68-71.
[4]潘莺,王善元.模拟熔喷纤网的孔径分布及其对过滤性能的影响[J].中国纺织大学学报, 2000, 26(5):73-77.
PAN Ying, Wang Shanyuan. Simulation of melt blown web and the effect of its poresize distribution on filtration performance[J]. Journal of China Textile University, 2000, 26(5):73-77.
[5]SIMMONDS G E, BOMBERGER J D, BRYNER M A. Designing nonwovens to meet pore size specifications[J]. Journal of Engineered Fibers and Fabrics,2007,2(1):155892500700200.
[6]金关秀,祝成炎.基于图像模拟和支持向量机的复合非织造布孔隙尺寸预测[J].纺织学报,2022,43(12):75-81.
JIN Guanxiu,ZHU Chengyan. Prediction of pore dimension in composite nonwovens based on image simulation and support vector machine[J]. Journal of Textile Research,2022,43(12):75-81.
[7]杨旭红.非织造材料（纤维网）形态结构的表征与分形模拟[D].苏州:苏州大学,2003:27-30.
YANG Xuhong. Expression and fractal simulation of morphologic structures of nonwovens[D]. Suzhou: Soochow University,2003:27-30.
[8]高贵.利用分形理论研究非织造布的结构及透通性[D].天津：天津工业大学,2005:38-45.
GAO Gui. Using fractal theory to study the structure and permeability of nonwoven fabrics[D]. Tianjin：Tianjin Polytechnic University, 2005:38-45.
[9]WANG J, SHI Q, AKANKWASA N T, et al. Establishment of the three-dimensional model of the nonwoven structure with fiber scale based on the GAN algorithm[J]. Textile Research Journal, 2022, 92(9/10): 1656-1665.
[10]AYDILEK A H，OGUZ S H,EDIL T B. Constriction size of geotextile filters[J]. Journal of Geotechnical and Geoenvironmental Engineering, 2005, 131(1): 28-38.
[11]CRESWELL A, WHITE T, DUMOULIN V, et al. Generative adversarial networks: An overview[J]. IEEE Signal Processing Magazine, 2018, 35(1): 53-65.
[12]GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11): 139-144.
[13]LIU H, GU X, SAMARAS D. Wasserstein GAN with quadratic transport cost[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul, Korea (South). IEEE, 2020: 4831-4840.
[14]ZHU J Y, PARK T, ISOLA P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy. IEEE, 2017: 2242-2251.
[15]ZHANG H, GOODFELLOW I, METAXAS D, et al. Self-attention generative adversarial networks[EB/OL]. 2018: arXiv: 1805.08318. https://arxiv.org/abs/1805.08318.
[16]KARRAS T, AILA T, LAINE S, et al. Progressive growing of GANs for improved quality, stability, and variation[EB/OL].2017: arXiv:1710.10196. https://arxiv.org/abs/1710.10196.
[17]SHORTEN C, KHOSHGOFTAAR T M. A survey on image data augmentation for deep learning[J]. Journal of Big Data, 2019, 6(1):1-48.
[18]RADFORD A, METZ L, CHINTALA S. Unsupervised representation learning with deep convolutional generative adversarial networks[EB/OL].2015: arXiv:1511.06434.https://arxiv.org/abs/1511.06434.
[19]GULRAJANI I, AHMED F, ARJOVSKY M, et al. Improved training of wasserstein gans[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems,2017:5769-5779.
[20]BERTHELOT D, SCHUMM T, METZ L. Began: Boundary equilibrium generative adversarial networks[EB/OL].2015: arXiv:1511.06434. https://arxiv.org/abs/1511.06434.
[21]张潇,贺晓亚.计算机图像技术在织物性能测试分析中的应用[J].国际纺织导报,2022,50(4):32-35.
ZHANG Xiao, He Xiaoya. Application of computer image technology in fabric properties testing and analysis[J]. Melliand China,2022,50(4):32-35.