%0 Journal Article
%T TrisimpleNet：基于三元损失和特征合成的缺陷检测网络<br>TrisimpleNet: Defect Detection Network Based on Triplet Loss and Feature Synthesis
%A 毛加平
%J Advances in Applied Mathematics
%P 2323-2330
%@ 2324-8009
%D 2024
%I Hans Publishing
%R 10.12677/aam.2024.135219
%X 无监督的缺陷检测模型往往需要合成样本，然而研究者大多聚焦于缺陷样本的生成，对正常样本的处理较少。这种偏向性导致了模型对正常样本的过拟合，影响了其在实际应用中的性能。受到之前研究中特征加噪生成缺陷特征的启发，同时观察到广泛使用的MVTec AD缺陷数据集中几种类别的数据集容易受到微小扰动影响而被误判为缺陷，为了解决正常样本受随机性干扰的问题，本研究通过新的思路提出了一个名为TrisimpleNet的模型。模型在生成缺陷样本的同时生成伪造的正常样本，从而更全面地模拟实际情况。此外，基于三元组损失的思想对损失函数进行了优化，使模型对正常样本的微小扰动具有更强的鲁棒性，从而降低了假阴性的概率。经实验模型在数据集上的取得了较好的表现，测试中几种类别的AUROC均得到提升。<br />
Unsupervised defect detection models often require synthetic samples; however, researchers tend to focus more on generating defect samples and pay less attention to normal samples. This bias leads to overfitting of the model on normal samples, affecting its performance in practical applications. Inspired by previous studies on generating defect features by adding noise to features, and observing that several categories of the widely used MVTecAD defect dataset are prone to being misclassified as defects due to minor disturbances, this study proposes a model named TrisimpleNet to address the issue of normal samples being affected by randomness. The model generates fake normal samples while generating defect samples, thus more comprehensively simulating realworld scenarios. In addition, the loss function is optimized based on the idea of triplet loss, making the model more robust to minor disturbances in normal samples, thereby reducing the probability of false negatives. Experimental results show that the model performs well on the dataset, with improvements in AUROC for several categories during testing.
%K 缺陷合成，正常特征合成，三元组损失<br>Defect Synthesis
%K Normal Feature Synthesis
%K Triplet Loss
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=88417