| 张桢桢， 张团善， 江小丽.基于改进UNet的织物缺陷分割方法[J].轻工机械,2025,43(2):77-83 |
| 基于改进UNet的织物缺陷分割方法 |
| Fabric Defects Segmentation Method Based on Improved Unet |
| |
| DOI：10.3969/j.issn.1005 2895.2025.02.010 |
| 中文关键词: 织物 疵点检测 卷积注意力模块 高效多尺度注意力机制 语义分割 深度学习 |
| 英文关键词:fabric defect detection CBAM（Convolutional Block Attention Module） EMA（Efficient Multi Scale Attention） semantic segmentation deep learning |
| 基金项目:国家自然科学基金项目（51735010）。 |
|
| 摘要点击次数: 20 |
| 全文下载次数: 11 |
| 中文摘要: |
| 针对UNet模型对机织物疵点进行语义分割时所存在的边界检测不完整、分割区域不联通等检测准确率不高的问题，课题组提出了一种基于改进UNet模型的织物缺陷分割算法。为了解决UNet模型在跳跃连接设计中采用直接拼接的方式易产生噪声点的问题，在特征层引入注意力模块（Convolutional Block Attention Module，CBAM）消除影响，在解码端的上采样部分引入了一种可跨空间学习的高效多尺度注意力机制（Efficient Multi Scale Attention，EMA），增强有效特征的权重，抑制冗余特征。实验结果表明：改进后的UNet模型准确率达到99.16%，平均交并比达到78.53%，相较于vgg16 UNet模型、ResNet50 UNet模型以及DeeplabV3+模型性能更优，具有一定的工业应用价值。 |
| 英文摘要: |
| In response to the problems of incomplete boundary detection and disconnected segmentation regions in the semantic segmentation of woven fabric defects using the UNet model, an improved UNet based fabric defect segmentation algorithm was proposed. To address the issue of generating noise points in the design of skip connections using direct cascading in UNet networks, a Convolutional Block Attention Module（CBAM）was introduced in the feature layer to eliminate the impact,and an Efficient Multi Scale Attention（EMA） mechanism that can learn across spaces was introduced in the upsampling part of the decoding end to enhance the weight of effective features and suppress redundant features. The experimental results show that the improved UNet network has an accuracy of 99.16% and an average intersection to union ratio of 78.53%. Compared with the vgg16 UNet model, ResNet50 UNet model, and DeeplabV3+model,it has better performance and certain industrial value. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
