Deep Learning Based Simple CNN Weld Defects Classification Using Optimization Technique

• Published in: Russian journal of nondestructive testing 2022-06, Vol.58 (6), p.499-509
• Main Authors: Kumaresan, Samuel, Aultrin, K. S. Jai, Kumar, S. S., Anand, M. Dev
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Characterization and Evaluation of Materials; Chemistry and Materials Science; Classification; Computer architecture; Deep learning; Heterogeneity; Image contrast; Image processing; Image quality; Machine learning; Materials Science; Nondestructive testing; Optimization; Structural Materials; Weld defects; X-Ray Methods
• ISSN: 1061-8309; 1608-3385
• DOI: 10.1134/S1061830922060109

Weld fault identification using X-ray pictures is a useful nondestructive testing technique. Traditionally, this job has relied on skilled human experts, while the extraction and classification of heterogeneity required their personal involvement. To overcome those challenges, various approaches are used, including machine learning (ML) and image processing technologies. Although the detection and categorization of low contrast and poor-quality images have been improved, the end result is still unsatisfactory. Unlike earlier ML-based research, this paper provides a new deep learning network-based classification approach. In this work base convolutional neural network architecture with optimization techniques was used to enhance the performance of the architecture to obtain a best performance result in simpler CNN architecture. In this work we have obtained overall classification accuracy of 89% using simple CNN with optimization technique.