Enhancement of Anomaly Detection Using Two-Stage Anomlay Segmentation Model

EasyChair Preprint 9172

Abstract

For the anomaly detection task, previously presented deep learning approaches suffer from one potential issue in the testing stage, the resultant output image has noise and missing anomaly area. To deal with this issue, we present a novel two-stage convolutional neural network (CNN) for anomaly detection. In the training stage, the first model is trained by inserting pseudo-anomalies, while the second model is trained by a superpixel technique which segments the image refined by the first model. The superpixel technique can recover partially visible anomaly patterns and suppress noise outside the recovered anomaly patches. We trained the proposed model using an industrial dataset MVTec and compared its performance with state-of-the-art pseudo-anomalous method [11]. Our method shows comparable pixel based percentage area under the receiver operating characteristic (%AUROC) of 96.0% which is only 1.3% less than the performance of DRAEM. However, our model uses four times less number of parameters.

Keyphrases: Convolutional Neural Network, anomaly detection, segmentation model

@booklet{EasyChair:9172,
  author    = {Rizwan Ali Shah and HyungWon Kim},
  title     = {Enhancement of Anomaly Detection Using  Two-Stage Anomlay Segmentation Model},
  howpublished = {EasyChair Preprint 9172},
  year      = {EasyChair, 2022}}