Journal
JOURNAL OF CULTURAL HERITAGE
Volume 62, Issue -, Pages 404-411Publisher
ELSEVIER FRANCE-EDITIONS SCIENTIFIQUES MEDICALES ELSEVIER
DOI: 10.1016/j.culher.2023.06.018
Keywords
Infrared thermography; Machine learning; Non-destructive testing; Cultural heritage; Painting
Ask authors/readers for more resources
Early defects in artworks can worsen if not detected in time, and infrared-based non-destructive testing is considered an effective method. However, current methods still rely on human judgment, leading to bias and inconsistency. Therefore, an improved YOLOX algorithm with CBAM mechanism is used to detect defects in artwork using thermal imaging data.
Early defects in artworks, which are unnoticeable by the naked eye, can evolve in severity if left unattended; therefore, their timely and accurate detection is of vital importance. Infrared-based non-destructive testing (NDT) methods are widely seen as an effective means to detect defects in artworks. However, most existing methods still rely on the human-based judgment of defects after processing the thermal images, which would inevitably lead to opinion bias and inconsistency. Deep-learning algorithms for computer vision are thought adequate to address the situation. On this point, an improved YOLOX algorithm, with the convolutional block attention module (CBAM) mechanism added to the network, is employed to improve the status quo. The method works with thermal imaging data under flash lamps for tempera paintings on canvas. A part of the famous The Birth of Venus by Botticelli ( & SIM;1485 ca.) was re-produced by a professional restorer, and three different defects were artificially introduced. To overcome the challenge pertaining to having too few defect samples, the dataset is expanded using finite-element simulation methods. The simulated data are used for training, whereas the replica is used for testing. Experimental results show that adding the CBAM mechanism after upsampling and downsampling of the path aggregation network can maximize network performance. The performance of the improved network herein proposed outperforms the benchmarks, with an average accuracy of 96.67%, a detection speed of 20.0 fps, and a total of 5.08 million parameters.& COPY; 2023 Consiglio Nazionale delle Ricerche (CNR). Published by Elsevier Masson SAS. All rights reserved.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available