| 高港,魏利胜,朱圣博.正负样本差异特征双径向融合的自监督缺陷检测方法[J].电子测量与仪器学报,2024,38(5):201-209 |
| 正负样本差异特征双径向融合的自监督缺陷检测方法 |
| Self-supervised defect detection based on biradial fusion of differentialfeatures between positive and negative samples |
| |
| DOI： |
| 中文关键词: 缺陷检测 合成负样本 CA PANet 加权损失 |
| 英文关键词:defect detection feature matching CA PANet weighted loss |
| 基金项目:安徽省教育厅自然科学研究重大基金资助项目(KJ2020ZD39)、安徽省检测技术与节能装置重点实验室开放基金项目(DTESD2020A02）资助 |
|
|
|
|
| 摘要点击次数: 181 |
| 全文下载次数: 10479 |
| 中文摘要: |
| 针对纹理图像表面划痕、裂纹等缺陷不规则、随机分布,导致缺陷检测准确率低的问题,研究一种基于正负样本差异特征双径向融合的自监督缺陷检测方法。首先,采用Otsu阈值分割提取图像前景信息,并以DTD数据集中的纹理图像或数据增强后的正样本叠加Perlin噪声,对正样本图像进行缺陷模拟以合成负样本;然后,利用正负样本经编码器输出的中间特征,计算均方误差进行特征匹配,结合坐标注意力(coordinate attention, CA)和双径向路径聚合网络(path aggregation network, PANet)加强匹配特征的信息融合;最后,将融合特征与编码器输出的低层和高层特征一同输入解码器,优化调整Focal、L1和Dice损失函数权重,实现对缺陷掩码更精准地预测。实验显示,所提模型在MVTec AD数据集纹理类别上的平均图像级、像素级AUROC分别达到了0.995、0.968,相较于其他缺陷检测模型,分类和分割准确率均有提升,表明所提方法在纹理缺陷检测方面的有效性。 |
| 英文摘要: |
| Aiming at the problem of irregular and random distribution of defects on the surface of texture images, such as scratches and cracks, which leads to low accuracy of defect detection, a self-supervised defect detection method based on the bi-radial fusion of positive and negative sample difference features is proposed. Firstly, Otsu threshold segmentation is used to extract image foreground information, and Perlin noise is superimposed on the data-enhanced positive samples or the texture images, from the DTD dataset, to simulate defects on the positive sample images and synthesize the negative samples. Then, the mean-square error is calculated for feature matching using the intermediate features output from the encoder, while the coordinate attention (CA) and path aggregation network (PANet) are combined to enhance the information fusion of the matched features. Finally, the fused features are input into the decoder together with the low-level and high-level features output from the encoder, and the weights of Focal, L1, and Dice loss functions are optimized and adjusted to realize the prediction of the defective masks more accurately. Experiments show that the average image level and pixel-level AUROC of the proposed model on the texture category of the MVTec AD dataset reaches 0.995 and 0.968, respectively, which improves the classification and segmentation accuracies compared with the other defect detection models, demonstrating the effectiveness of the proposed method in texture defect detection. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
