Anomaly Detection using Deep Learning based Image Completion

  • 2018-11-16 15:36:28
  • Matthias Haselmann, Dieter P. Gruber, Paul Tabatabai
  • 5

Abstract

Automated surface inspection is an important task in many manufacturingindustries and often requires machine learning driven solutions. Supervisedapproaches, however, can be challenging, since it is often difficult to obtainlarge amounts of labeled training data. In this work, we instead performone-class unsupervised learning on fault-free samples by training a deepconvolutional neural network to complete images whose center regions are cutout. Since the network is trained exclusively on fault-free data, it completesthe image patches with a fault-free version of the missing image region. Thepixel-wise reconstruction error within the cut out region is an anomaly imagewhich can be used for anomaly detection. Results on surface images of decoratedplastic parts demonstrate that this approach is suitable for detection ofvisible anomalies and moreover surpasses all other tested methods.

 

Quick Read (beta)

loading the full paper ...