Anomaly detection for vision-based inspection

Chew, Yan Zhe (2022) Anomaly detection for vision-based inspection. Final Year Project, UTAR.

Preview

PDF Download (3392Kb) | Preview

Abstract

The usage of Artificial Intelligence (AI) techniques in vision-based anomaly detection is gaining traction within the manufacturing industry for quality inspection purposes. The implementation of visual inspection in a production line can effectively detect defective products, while saving time and cost by increasing the efficiency through process automation. There are two approaches to visual inspection: the conventional approach which uses image processing techniques and the modern AI-based approach through deep learning. This study aims to implement visual inspection systems using both approaches and determine the suitability of each approach for visual anomaly detection. Tests were conducted on the MVTec Anomaly Detection dataset, which is a dataset for benchmarking anomaly detection methods with a focus on industrial inspection. The algorithms that were used for the conventional approach are Template Matching, Structural Similarity Index, Scale Invariant Feature-Transform and Oriented FAST and Rotated BRIEF, whereas the deep learning models that were used for the modern approach are Patch Distribution Modelling (PaDiM) and PatchCore. The results demonstrate that conventional approaches are not suitable for anomaly detection, whereas modern AI approaches are able to detect anomalies and segmentate the area with a high degree of accuracy. The average Image Area Under the Receiver-Operator Characteristic Curve (AUROC) and Image F1 Score for the PaDiM model is 0.9151 and 0.9033, whereas for the PatchCore model the scores are 0.9993 and 0.9979 respectively. Nevertheless, there are still some instances where AI models will fail to perform as intended, but generally the performance is good. Future work regarding visual anomaly detection should be focused on modern AI-based approach, but classical methods can be applied at other suitable areas where the effectiveness to complexity trade-off is warranted.

Actions (login required)