Convolutional Neural Network-Based Classification System for Weld Radiographic Test Images

Wook Jin Cho; Haifeng Zhou; Jungchul Choi; Hee-Chang Lim

doi:10.5750/ijme.v168iA4(S).1845

PDF (GBP 29.99)

Published: Oct 10, 2026

DOI: https://doi.org/10.5750/ijme.v168iA4(S).1845

Keywords:

Convolutional neural network Image classification Radiographic test

Wook Jin Cho

Pusan National University, South Korea

Haifeng Zhou

Pusan National University, South Korea

Jungchul Choi

Mokpo National University, South Korea

Hee-Chang Lim

Pusan National University, South Korea

Abstract

Steel pipes are widely applied in shipbuilding and marine engineering. Welded steel pipes undertake heavy loads in
relevant structures, so non-destructive testing of welded joints is necessary to ensure integrity. Traditional radiographic testing is time-consuming and complex, relying on manual operation and evaluation by qualified inspectors, which easily causes human errors. This study adopts EfficientNet with transfer learning and fine-tuning to classify the quality of welded steel pipe radiographic images as qualified or unqualified under limited data. The test accuracy reaches 96.875%, showing excellent performance. This technology can be extended to RT-related industries such as steel pipe manufacturing, shipbuilding, automotive, and aerospace.

Issue

Vol. 168 No. A4(S) (2026): Special Issue: International Conference of Advance Research and Innovation

Section

Articles

Submissions

To make a Submission to IJME, click here

Subscriptions

To Subscribe to IJME, Click Here for Subscription Options

Information

Join the Royal Institution of Naval Architects

The University of Buckingham Press is delighted to partner with RINA on the publication of IJME. RINA is a global membership body for the maritime industry and has Members in 140 countries. To see the options of becoming a Member at RINA, click here

UBP Journal Policies

Peer Review

Once articles are accepted to peer-review, all articles must be blind peer reviewed by at least two reviewers. Each reviewer will submit their review including comments and a recommendation when to accept for publication, reject or to request revisions to the articles.

Respect for Participants:

All Journals will treat all individuals involved in research with dignity and respect..

Informed Consent:

Authors must obtain and document informed consent from participants, ensuring they understand the research purpose, procedures, and potential risks.

Data Protection and Confidentiality:

See University of Buckingham’s Press’ Data Protection Policy.

Conflict of Interest Disclosure:

Authors are required to disclose any potential conflicts of interest, including financial or personal relationships that might influence their research or its interpretation.

Authorship and Attribution:

All individuals who made substantial contributions to the research must be acknowledged and those listed as authors must have approved the final text.

Data Integrity and Reporting:

Authors must accurately present their research findings, avoid fabrication or falsification of data, and provide sufficient detail for replication.

Plagiarism:

Journals should confirm they are using established plagiarism software, ensuring that submitted work is original and appropriately cites sources.

Ethical Approval:

For research involving human or animal subjects, it is required that confirmation that ethical approval was obtained from relevant institutional review boards or ethics committees.

Scientific Misconduct:

Any suspected scientific misconduct, such as fabrication, falsification, or plagiarism, must be immediately reported to the Publisher by emailing info@unibuckinghampress.com

Article Sidebar

Main Article Content

Abstract

Article Details

Most read articles by the same author(s)