| 马浩然， 张团善， 王峰， 赵浩铭.基于语义生成与语义分割的机织物缺陷检测方法[J].轻工机械,2023,41(1):66-73 |
| 基于语义生成与语义分割的机织物缺陷检测方法 |
| Woven Fabric Defect Detection Method Based on Semantic Generation and Semantic Segmentation |
| |
| DOI：10.3969/j.issn.1005 2895.2023.01.011 |
| 中文关键词: 机织物缺陷检测 语义分割 语义生成网络 尺寸自适应Dice损失函数 BEGAN 免标注 |
| 英文关键词:woven fabric defect detection semantic segmentation semantic generation network size adaptive Dice loss function BEGAN annotation free |
| 基金项目:西安市现代智能纺织设备重点实验室资助项目(2019220614SYS021CG043)。 |
|
| 摘要点击次数: 126 |
| 全文下载次数: 108 |
| 中文摘要: |
| 针对织物疵点的语义分割任务中因数据集规模限制，而导致网络出现的严重过拟合问题，课题组提出了针对织物的语义生成网络。语义生成网络使用随机产生的语义标签生成对应的织物缺陷图像，相较于传统数据增强方法，语义生成可生成全新图像，更贴近实际缺陷分布，并且可通过判别器对生成图像进行筛选；课题组将语义生成的图像作为语义分割网络的输入，相应的随机语义标签作为目标，免去标注过程，扩充语义分割网络的训练样本，提升网络性能；对于语义分割网络，提出尺寸自适应Dice损失函数，解决样本不平衡问题，提升网络对小尺寸的检测能力。实验结果表明：尺寸自适应Dice损失函数使得模型精度提高11.1%，使用BEGAN扩充的数据集相较于传统方法扩充的数据集训练得到的模型精度提高7.4%。 |
| 英文摘要: |
| Aiming at the severe overfitting problems caused by the limitation of data set size in the semantic segmentation task of fabric defects, a semantic generation network for fabric which uses randomly generated semantic labels to generate corresponding fabric defect images was proposed. Compared to traditional data enhancement methods, semantic generation can generate brand new images that are closer to the actual defect distribution, and the generated images can be filtered by discriminators. The semantically generated images were used as the input of the semantic segmentation network, and the corresponding random semantic labels were used as the target, so as to eliminate the labeling process, expand the training samples of the semantic segmentation network and improve the network performance. For the semantic segmentation network, the size adaptive Dice loss function was proposed to solve the sample imbalance problem and improve the detection ability of the network for small sizes. The experimental results show that the size adaptive Dice loss function improves the model accuracy by 11.1%, and the accuracy of the model obtained by training with the BEGAN expanded dataset is improved by 7.4% compared with that of the dataset expanded by the traditional method. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
