SIFT and SURF based feature extraction for the anomaly detection

TL;DR

This paper demonstrates that SIFT and SURF algorithms effectively extract features for anomaly detection, enabling classifiers to achieve around 89% accuracy in semi-supervised and one-class settings, with publicly available dataset and code.

Contribution

It introduces a method using SIFT and SURF for feature extraction in anomaly detection, applicable in semi-supervised and one-class classification scenarios.

Findings

SIFT and SURF can be used as effective feature extractors for anomaly detection.

Classifiers trained on these features achieve approximately 89% accuracy.

Performance of one-class classifiers is comparable to semi-supervised classifiers.

Abstract

In this paper, we suggest a way, how to use SIFT and SURF algorithms to extract the image features for anomaly detection. We use those feature vectors to train various classifiers on a real-world dataset in the semi -supervised (with a small number of faulty samples) manner with a large number of classifiers and in the one-class (with no faulty samples) manner using the SVDD and SVM classifier. We prove, that the SIFT and SURF algorithms could be used as feature extractors, that they could be used to train a semi-supervised and one-class classifier with an accuracy around 89\% and that the performance of the one-class classifier could be comparable to the semi-supervised one. We also made our dataset and source code publicly available.