Patch SVDD(support vector data description)-based channel attention embedding and improvement of classifier

Abstract

Some anomaly detection methods are based on CNN to fuse spatial and channel-wise information together within local receptive fields. However, the correlation between feature channels has not been fully utilized. Channel attention has been shown to model the interdependence between convolution feature channels and improve network representation. It is possible to introduce channel attention into anomaly detection. We attempt to directly embed the SE(Squeeze and Excitation) module into the convolutional layer but reduced anomaly detection performance. Therefore, we propose a lightweight channel attention module C-SE(Current Squeeze and Excitation) suitable for anomaly detection. C-SE module not only improves the representation ability of depth convolutional neural network but also has a significant effect on texture anomaly detection. C-SE module body is constructed by average pooling and maximum pooling branches, which ensure that local salient features of the image are not lost. Then reduce the negative impact of feature calibration through a long connection. In addition, the improvement of classifier plays an important role. Experimental results have shown that the proposed method outperforms the Patch SVDD methods by 3% in image-level AUROC and 0.7% in pixel-level AUROC on the MVTec benchmark. The higher AUROC score and the faster rate of convergence prove the effectiveness of the method.