| 摘 要: 针对工业缺陷检测中的缺陷样本稀缺及受限于二维RGB图像(即二维红绿蓝色彩图像)无法表征几何缺陷特征的问题,提出了一种基于多模态记忆库(Memory Bank)的三维缺陷检测方法。该方法通过对比学习对齐匹配不同模态的特征;使用缺陷过滤器滤除缺陷样本中的大量噪声,并构建高质量的多模态记忆库;利用弹性特征分类器检测缺陷。实验结果显示,在每一种缺陷仅有4张训练图像的极端条件下,相较于最优的二维缺陷检测方法PatchCore在公开数据集上仅能达到38.72%的准确率,提出的三维缺陷检测方法则能达到43.14%的准确率,更适用于缺陷样本稀缺的工业场景。 |
| 关键词: 多模态;记忆库;三维缺陷检测;计算机视觉 |
|
中图分类号: TP391.4
文献标识码: A
|
|
| 3D Defect Detection Method Based on Multimodal Memory Bank |
|
WANG Qinyang, CHEN Jing
|
(School of Computer Science, Hangzhou Dianzi University, Hangzhou 310018, China)
wqy1301670302@163.com; cj@hdu.edu.cn
|
| Abstract: To address the challenges of scarce defect samples in industrial defect detection and the limitations of 2D RGB images (i.e., two-dimensional red-green-blue color images) in characterizing geometric defect features, this paper proposes a 3D defect detection method based on a multimodal memory bank. The method aligns and matches features across different modalities through contrastive learning and employs a defect filter to remove significant noise from defect samples. Additionally, a high-quality multimodal memory bank is constructed, and an elastic feature classifier is utilized for defect detection. Experimental results demonstrate that, under extreme conditions with only four training images per defect, the proposed 3D defect detection method achieves an accuracy of 43.14% , outperforming the state-o-f the-art 2D defect detection method PatchCore, which achieves only 38.72% accuracy on public datasets.The proposed method proves to be suitable for industrial scenarios with scarce defect samples. |
| Keywords: multimodal; memory bank; 3D defect detection; computer vision |
