Integrated CNN for Specified Object Filtering and Abnormal Motion Detection for Smart Factory 


Vol. 48,  No. 8, pp. 1001-1011, Aug.  2023
10.7840/kics.2023.48.8.1001


PDF
  Abstract

Autoencoders[1,2,3], which reconstruct the image at the output only to an image similar to their training images, have been widely used in anomaly detection networks[4]. They has the advantage of extracting abnormal motions in the input image because it has the characteristic of regenerating the output image by limiting the input image to data similar to the previously learned image data[4]. However, the anomaly detection method based on an autoencoder has a disadvantage in that it cannot correctly detect a situation in which there is an object to be judged normal at an irregular location in the input image. To address this problem, we propose an integrated CNN with both an autoencoder and an object segmentation heads[5]. We introduce a segmentation-based object masking technique that can exclude normal areas of the input image and reconstructed output image. It can improve the accuracy of detecting anomalies in images where the normal objects are positioned in an irregular location in the image. In addition, we propose an automatic labeling technique which utilizes an autoencoder of the integrated CNN to add pseudo labels[6] to the images containing normal objects and so effectively conducts unsupervised learning[7]. We applied the proposed integrated CNN model to a video dataset obtained from a smart factory. Our unsupervised learning technique demonstrated a mask mAP of 96.82% for the segmentation model when tested using hand-labeled ground-truth data. In addition, the proposed integrated CNN model improves the overall anomaly detection accuracy by 15.90% by increasing the accuracy of detecting normal objects under the situation where both normal and abnormal objects coexist, which the existing autoencoder was not able to detect.

  Statistics
Cumulative Counts from November, 2022
Multiple requests among the same browser session are counted as one view. If you mouse over a chart, the values of data points will be shown.


  Related Articles
  Cite this article

[IEEE Style]

S. Hong and H. Kim, "Integrated CNN for Specified Object Filtering and Abnormal Motion Detection for Smart Factory," The Journal of Korean Institute of Communications and Information Sciences, vol. 48, no. 8, pp. 1001-1011, 2023. DOI: 10.7840/kics.2023.48.8.1001.

[ACM Style]

Sang-wook Hong and Hyung-won Kim. 2023. Integrated CNN for Specified Object Filtering and Abnormal Motion Detection for Smart Factory. The Journal of Korean Institute of Communications and Information Sciences, 48, 8, (2023), 1001-1011. DOI: 10.7840/kics.2023.48.8.1001.

[KICS Style]

Sang-wook Hong and Hyung-won Kim, "Integrated CNN for Specified Object Filtering and Abnormal Motion Detection for Smart Factory," The Journal of Korean Institute of Communications and Information Sciences, vol. 48, no. 8, pp. 1001-1011, 8. 2023. (https://doi.org/10.7840/kics.2023.48.8.1001)
Vol. 48, No. 8 Index

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)






pISSN : 1226 - 4717
eISSN : 2287 - 3880