International Journal on Advanced Science, Engineering and Information Technology, Vol. 13 (2023) No. 4, pages: 1480-1485, DOI:10.18517/ijaseit.13.4.19020

PatchCore-based Anomaly Detection using Major Object Segmentation

Gyu-Il Kim, Kyungyong Chung


Cameras utilized for product defect detection in the production line of the manufacturing process create noise due to environmental changes such as camera angle and direction of light. This causes a lack of manufacturing process data and reduces the efficiency of anomaly detection. Therefore, it is necessary to produce a method that detects defects occurring in the production line and guarantees product quality and safety using anomaly detection technology combined with artificial intelligence. Therefore, this thesis proposes PatchCore-based anomaly detection using major object segmentation. The proposed method pre-processes product packaging data by using Green Channel thresholding, Major Connected Component Selection, Extracting Outer Contour, and FloodFill with Centroid. As for the pre-processed data main objects are masked, and the image data is segmented. Through PatchCore model, normality and anomaly detection results are binarily classified. In the performance evaluation, the accuracy is compared between the pre-existing anomaly detection method and the proposed method through the pre-/post-preprocessing data, and high performance is proven. The conventional method showed an accuracy of 0.7684, while our approach achieved an accuracy of 0.9784. Additionally, among the CNN models, VGG19 demonstrated an accuracy of 0.5833, and EfficientNet80 showed an accuracy of 0.7, both of which were lower than our method's accuracy. Therefore, even a small data set shows strong performance through the proposed method. The proposed method is expected to be utilized as an effective defect detection model in diverse fields.


Anomaly detection; deep learning; machine learning; artificial intelligence; image segmentation

Viewed: 279 times (since abstract online)

cite this paper     download