| 摘 要: 织物瑕疵种类繁多且获取困难,导致织物瑕疵检测具有一定的挑战性。为解决此难题,提出了一种基于改进CycleGAN模型的织物瑕疵图像生成方法,旨在丰富织物数据集。引入U-Net并对其特征提取模块进行优化,设计一种并行扩张双向注意力结构,以有效提取图像纹理和边缘特征信息。在快速连接中,加入混合注意力结构以有效过滤冗余信息,并设计深度残差结构增强模型表达能力和网络深度。实验结果表明,该模型在FID、PSNR及LPIPS等关键指标上均表现出显著优势,充分验证了其优越性。 |
| 关键词: CycleGAN 织物瑕疵生成 生成对抗网络 无监督学习 |
|
中图分类号: TP391
文献标识码: A
|
|
| Unsupervised Fabric Defect Generation Based on Improved CycleGAN |
|
ZHANG Jinfeng, ZHANG Jiehao, XIANG Zhong
|
(School of Mechanical Engineering, Zhejiang Sc-i Tech University, Hangzhou 310018, China)
zjf 971120@163.com; zszhangjh1999@163.com; xz@zstu.edu.cn
|
| Abstract: The diversity and scarcity of fabric defects pose significant challenges to defect detection in textile manufacturing. To address this issue, this paper proposes an improved CycleGAN-based defect generation method to enrich fabric datasets. We introduce a U-Net architecture with optimized feature extraction modules and design a parallel dilated bidirectional attention structure to effectively capture texture and edge features. Within the skip connections, a hybrid attention mechanism is incorporated to filter redundant information, while a deep residual structure enhances the model’s expressive capacity and network depth. Experimental results demonstrate the model’s superior performance across key metrics (FID, PSNR, and LPIPS), validating its effectiveness. |
| Keywords: CycleGAN fabric defect generation generative adversarial networks unsupervised learning |
