基于度量学习的小样本零器件表面缺陷检测

Few-shot parts surface defect detection based on the metric learning

针对零器件表面缺陷检测时存在缺陷样本少、缺陷目标尺寸大小不一和易发生几何形变等问题,提出一种基于度量学习的小样本零器件表面缺陷检测模型。该模型首先将特征金字塔网络中传统卷积改进为动态卷积,并加上区域建议网络对小样本缺陷进行特征提取和边框定位;然后在大型数据集MS COCO上进行预训练,将训练好的模型结构参数迁移到具有少量缺陷样本的检测中;最后建立基于度量学习的多模态网络结构实现小样本零器件表面缺陷检测。实验表明,所提方法在ImageNet LOC公共数据集上与其他模型相比性能更优,5类5样本下均值平均精度为70.43%;在所建立的零器件表面缺陷数据集上,3类5样本的均值平均精度最高可达35.76%,相比RepMet模型性能最大可多提升近70%。

The detection of parts surface defects faces the problems of few defect samples, different defect target sizes and easy geometric deformation. To address these issues, this paper proposes a few-shot parts surface defect detection model based on the metric learning. First, the traditional convolution in the feature pyramid network is improved to deformable convolution, and the region proposal network is used to extract features and frame positioning from few-shot defects. Then, it is pre-trained by using the large data set MS COCO. The trained model structure parameters are transferred to the detection of a small number of defect samples. Finally, the multi-modal network structure based on the metric learning is established to realize few-shot parts surface defect detection. Experimental results show that the proposed method performs better than other models on the ImageNet LOC public data set. The mean average precision under the 5-way 5-shot is 70.43%. For the established parts surface defect data set, the mean average precision under the 3-way 5-shot is up to 35.76%. On the parts surface defect data set, the feasibility and effectiveness of the model are evaluated. The mean average precision under the 3-way 5-shot is up to 35.76%. Compared with the RepMet model, the performance improvement can be increased by nearly 70%.