融合自注意力与缺陷凸显的缺陷检测方法

Defect detection method integrating self-attention and highlighting of defects

针对无监督缺陷检测中重建网络在抑制异常重建的同时无法保留正常区域细节信息的问题,提出了一种融合自注意力与缺陷凸显的缺陷检测方法。首先,在重建网络中引入离散小波变换(DWT)进行下采样,并使用离散小波逆变换(IDWT)进行上采样。相较于传统重建网络,这种方法能减少细节信息的丢失,并对特征进行频率分解。同时,在跳连接中加入自注意力模块对特征重新编码,使其重点关注正常区域的细节信息。此外,设计了一个缺陷区域凸显模块,利用正常样本特征构建特征库,将从测试图像提取的特征与特征库中特征对比得到异常图,将异常图与重建差值图相结合来改善缺陷定位结果。在工业缺陷检测数据集MVTec AD上进行测试,在图像级AUROC上取得了99.3%的结果,同时在像素级AUROC上取得了98.3%的结果,在无监督缺陷检测中具有较高的检测精度和鲁棒性。

To address the issue of reconstruction networks in unsupervised defect detection failing to preserve detailed information of normal regions while simultaneously suppressing abnormal reconstructions,a defect detection method that combines self-attention and defect highlighting is proposed.First,the discrete wavelet transform(DWT)is introduced in the reconstruction network for downsampling,and the inverse discrete wavelet transform(IDWT)is used for upsampling.Compared to traditional reconstruction networks,this method reduces the loss of detail information and performs frequency decomposition on features.Then self-attention modules are added into the skip connections to re-encode the features,enabling the features to focus on the details of the normal region.Additionally,a defect region highlighting module is designed,which utilizes features from normal samples to construct a feature library.By comparing the features extracted from the test image with the features in the library,an abnormal map is obtained.Finally,the abnormal map is combined with the reconstruction residual map to improve the results of defect localization.The proposed method is tested on the industrial defect detection dataset MVTec AD and achieved 99.3%area under the receiver operating characteristic curve(AUROC)at the image level and 98.3%at the pixel level,demonstrating high detection accuracy and robustness in unsupervised defect detection.