Title: Product surface defect detection algorithm based on transfer learning and DL

Authors: Xitao Sun; Shuo Xue

Addresses: School of Big Data and Artificial Intelligence, Anhui Xinhua University, Hefei, 230088, China ' School of Electrical and Information Engineering, Anhui University of Technology, Maanshan, 243002, China

Abstract: This study addresses the critical challenge of enhancing surface defect detection accuracy in industrial manufacturing through an optimised deep learning framework. We propose a hybrid model integrating convolutional block attention module (CBAM) and embedded inverse residual block (EIRB) into the U-Net architecture, combined with meta-transfer learning for structural optimisation. The enhanced network demonstrates superior performance: achieving 93.6% classification accuracy, 81% recall rate, and 8.98-second average detection time. Cross-part testing on six industrial components shows 92.5% accuracy for gears and 90.3% for valves. Notably, under varying lighting conditions, it maintains the highest F1-score of 0.89 compared to conventional models. This approach balances computational efficiency with detection robustness, making it suitable for real-time industrial applications requiring high precision and adaptability. The results confirm that our method provides both technical superiority and practical feasibility for advanced defect detection systems in manufacturing environments.

Keywords: industrial products; surface defect detection; transfer learning; meta transfer learning; MTL; U-Net.

DOI: 10.1504/IJSCC.2026.150319

International Journal of Systems, Control and Communications, 2026 Vol.17 No.1, pp.88 - 105

Received: 27 Feb 2025
Accepted: 21 May 2025
Published online: 09 Dec 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article