基于跨模态注意力聚合与自适应原型生成的小样本缺陷分割
CSTR:
作者:
作者单位:

东北大学 信息科学与工程学院,沈阳 110819

作者简介:

通讯作者:

E-mail: zhangyunzhou@mail.neu.edu.cn.

中图分类号:

TP391.41

基金项目:

国家自然科学基金项目(61973066);辽宁省科技计划重大项目子课题(2021JH1/10400049).


Few-shot defect segmentation based on cross-modal attention aggregation and adaptive prototype generation network
Author:
Affiliation:

College of Information Science and Engineering, Northeastern University, Shenyang 110819, China

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    基于深度学习的缺陷分割技术对于保证生产效率和改善产品质量至关重要.然而,在实际应用中有许多领域无法收集大规模的缺陷样本,导致传统缺陷检测方法性能急剧下降.此外,缺陷区域存在尺寸小、纹理信息弱以及与无缺陷区域对比不明显的问题,进一步阻碍了视觉缺陷检测技术的实际应用.对此,提出一种基于视觉与点云的多模态小样本缺陷分割方法,通过采用跨模态注意力聚合RGB语义信息和点云结构信息,实现两种模态的高效融合;然后,结合多模态特征与支持掩码生成基本前景原型、自适应背景原型和遗忘补偿原型,提升支持原型的表征能力;进而,根据相似性动态地匹配原型与查询特征,并在特征丰富化后完成对未见过物体缺陷的有效分割.在Defect-3i和Mvtec 3D-2i两个小样本缺陷分割数据集上的实验表明,所提出算法在单样本(1-shot)和五样本(5-shot)两种设置中的平均交并比(mIoU)分别超过其他先进小样本算法0.11%和0.20%、5.23%和5.10%,验证了所提出小样本架构的合理性与多模态网络的先进性.

    Abstract:

    Defect segmentation technology based on deep learning is crucial to ensure production efficiency and improve product quality. However, there are many areas in which large-scale defect samples cannot be collected in applications, resulting in a sharp decline in the performance of traditional detection methods. In addition, defect regions suffer from small size, weak texture information, and inconspicuous contrast with non-defect regions, which hinder the application of visual detection techniques. This paper proposes a multi-modal few-shot defect segmentation method based on vision and point cloud. Cross-modal attention is used to aggregate RGB semantic information and point cloud structure information to achieve efficient fusion of the two modalities. Then, basic foreground prototypes, adaptive background prototypes and forgetting compensation prototypes are generated by combining multi-modal features and masks to improve representation ability, dynamically match prototypes and query features according to the similarity, and complete effective segmentation of unseen object defects after feature enrichment. Experiments on two few-shot defect segmentation datasets, Defect-3i and Mvtec 3D-2i, show that the mean Intersection-over-Union(mIoU) in 1-shot and 5-shot settings exceeds other advanced algorithms by 0.11% and 0.20%, 5.23% and 5.10%, respectively, verifying the rationality of the proposed few-shot architecture and the advancement of the multi-modal network.

    参考文献
    相似文献
    引证文献
引用本文

刘世同,张云洲,单德兴,等.基于跨模态注意力聚合与自适应原型生成的小样本缺陷分割[J].控制与决策,2024,39(11):3655-3663

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2024-09-20
  • 出版日期: 2024-11-20
文章二维码