Buddhist Amulet Recognition by Using ResNet50
Issued Date
2022-12-01
Resource Type
eISSN
29850290
Scopus ID
2-s2.0-105005578887
Journal Title
Science Essence Journal
Volume
38
Issue
2
Start Page
1
End Page
14
Rights Holder(s)
SCOPUS
Bibliographic Citation
Science Essence Journal Vol.38 No.2 (2022) , 1-14
Suggested Citation
Pornpanomchai C., Pornpanomchai V. Buddhist Amulet Recognition by Using ResNet50. Science Essence Journal Vol.38 No.2 (2022) , 1-14. 14. Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/110396
Title
Buddhist Amulet Recognition by Using ResNet50
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
The objective of this research is to develop a computer system which can recognize Buddhist amulet images. The system is called “Buddhist amulet recognition system (BARS)”. BARS consists of four main modules, namely: 1) dataset training, 2) image acquisition, 3) ResNet50 classification and 4) result presentation. The system dataset consists of 3,248 images belonging to 203 amulet types, with 16 images per type. The system analyzed both metal & clay amulets, which consisted of 146 metal amulets and 57 clay ones. BARS employed the pre-training convolutional neural network (CNN) called “ResNet50” in MATLAB for recognizing Buddhist amulets. The accuracy, sensitivity, specificity and precision rates for the training dataset of BARS are 0.9998, 0.9879, 0.9999 and 0.9879, respectively. The system also conducted cross-validation on an untrained dataset, which has accuracy, sensitivity, specificity and precision rates of 0.9999, 0.9541, 0.9999 and 0.9541, respectively. The average training time is 3,183.2 seconds and the average access time is 1.34 second per image. Finally, this research compares the accuracy of ResNet18, ResNet50 and ResNet101, with the same amulet dataset.