Anomaly detection is the identification of any event that falls outside what is considered ‘acceptable behaviour’. This work investigates anomaly detection for automated visual inspection in the context of industry automation (‘Industry 4.0’). For this task we propose a machine vision procedure based on visual feature extraction and one-class k nearest neighbours classification. The method requires only samples of normal (non-defective) instances for the training step. We benchmarked our approach using seven traditional (‘hand-designed’) colour texture descriptors and five pre-trained convolutional neural networks (CNN) ‘off-theshelf’. Experimenting on nine image datasets from seven classes of materials (carpet, concrete, fabric, layered fused filament, leather, paper and wood), each containing normal and abnormal samples, we found overall accuracy in the range 82.0%–90.2%. Convolutional networks off-theshelf performed generally better than the traditional methods, although – interestingly – this was not true for all the datasets considered. No visual descriptor clearly emerged as the all-purpose best option.

A benchmark of traditional visual descriptors and convolutional networks ‘Off-the-Shelf’ for anomaly detection

Francesco Bianconi
;
Paolo Conti;Elisabetta Maria Zanetti;Giulia Pascoletti
2022

Abstract

Anomaly detection is the identification of any event that falls outside what is considered ‘acceptable behaviour’. This work investigates anomaly detection for automated visual inspection in the context of industry automation (‘Industry 4.0’). For this task we propose a machine vision procedure based on visual feature extraction and one-class k nearest neighbours classification. The method requires only samples of normal (non-defective) instances for the training step. We benchmarked our approach using seven traditional (‘hand-designed’) colour texture descriptors and five pre-trained convolutional neural networks (CNN) ‘off-theshelf’. Experimenting on nine image datasets from seven classes of materials (carpet, concrete, fabric, layered fused filament, leather, paper and wood), each containing normal and abnormal samples, we found overall accuracy in the range 82.0%–90.2%. Convolutional networks off-theshelf performed generally better than the traditional methods, although – interestingly – this was not true for all the datasets considered. No visual descriptor clearly emerged as the all-purpose best option.
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11391/1532774
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact