| 摘 要: 纺织工业验布环节自动检测是大势所趋,受工厂实际生产状况导致的异类瑕疵样本不均问题一直限制织物瑕疵检测算法的泛化应用。因此,提出基于改进StyleGAN2的织物瑕疵图像生成算法。在原算法架构上引入多尺度特征融合模块,帮助网络学习到关键特征;提出混合注意力机制,有效提取纹理和边缘特征信息;提出特征匹配和风格损失函数,增加细节特征的多样性和生成图像的保真性。所生成的织物瑕疵图像相比其他网络模型的图像评价指标FID(Fréchet Inception Distance)、峰值信噪比PSNR(Peak Signal-to-NoiseRatio)和感知图像损失LPIPS(Learned Perceptral Image Patch Similarity)分别至少提升36.35%、7.00%、62.50%,因此,进行图像生成可实现数据集的样本增强,进而提升检测算法的整体泛化性应用。 |
| 关键词: StyleGAN2 图像生成 织物瑕疵 生成对抗网络 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 江苏省自然科学基金项目(BK20171303) |
|
| Fabric Defect Image Generation Algorithm Based on Improved StyleGAN2 |
|
NIU Kui1, XIANG Zhong2
|
(1.School of Information Science and Engineering, Zhejiang Sc-i Tech University, Hangzhou 310018, China;
2.School of Mechanical Engineering, Zhejiang Sc-i Tech University, Hangzhou 310018, China)
nk13783066277@163.com; xz@zstu.edu.cn
|
| Abstract: Automated detection in fabric inspection has become an inevitable trend in the textile industry.However, the uneven distribution of heterogeneous defect samples caused by actual production conditions consistently limits the generalization of fabric defect detection algorithms. To address this, this paper proposes a fabric defect image generation algorithm based on improved StyleGAN2. The approach introduces a mult-i scale feature fusion module into the original architecture to enhance the network’s ability to learn key features; a hybrid attention mechanism is proposed to effectively extract texture and edge features; feature matching and style loss functions are designed to increase the diversity of detailed features and the fidelity of generated images. Compared with other network models, the generated fabric defect images achieve improvements of at least 36.35% , 7.00% , and 62.50% in evaluation metrics Fréchet Inception Distance, Peak Signa-l to-Noise Ratio, and Learned Perceptral Image Patch Similarity, respectively.Such image generation enables dataset augmentation and thereby enhances the overall generalization of detection algorithms. |
| Keywords: StyleGAN2 image generation fabric defects generative adversarial network |
