Hybrid twin attention based convolutional stacked sparse autoencoder for classification of defected weld images
Welding is an essential joining process in industrial manufacturing. Many deep learning models are introduced to detect welding errors. However, with a shortage of training data samples, most existing models take longer and are less accurate because of limited learning ability and increased computat...
Saved in:
| Published in: | Computers & electrical engineering Vol. 124; p. 110328 |
|---|---|
| Main Authors: | , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Ltd
01.05.2025
|
| Subjects: | |
| ISSN: | 0045-7906 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Welding is an essential joining process in industrial manufacturing. Many deep learning models are introduced to detect welding errors. However, with a shortage of training data samples, most existing models take longer and are less accurate because of limited learning ability and increased computational complexity problems. To address issues with existing methods, this research presents an efficient deep learning model for accurately classifying multiple welding flaws in minimal time. The most crucial steps that are carried out in the proposed welding error detection framework are pre-processing feature extraction and classification. Initially, the input images are collected from the welding defects dataset. To increase the quality of the obtained raw input images, different pre-processing techniques, such as image scaling, image denoising, and image enhancement are applied. After pre-processing, feature extraction is carried out with the help of the discrete wavelet transform (DWT) and the grey-level run length matrix (GLRLM), which helps to reduce the complexity problems. Finally, a Hybrid twin attention-based Convolutional Stacked Sparse Autoencoder (HAT_CS2E) is used to classify multiple weld defects accurately from the given images. The proposed model combines a convolutional neural network (CNN) and a stacked sparse autoencoder network. The integration of these networks helps to learn more spatial and local features that generate high quality feature maps and produce accurate classification outcomes. For simulation, the Welding Defects Dataset is utilized, and several existing approaches are compared with the proposed model in terms of accuracy, precision, recall, F1-score, and calculation time. The proposed model attained an accuracy of 97.01 %, precision of 96.98 %, recall of 95.76 %, F1-score of 95.12 %, and computation time of 0.021 s by altering frame level welding defect recognition. Also, the proposed model achieved superior results in pixel level welding defect detection process compared with existing studies in terms of accuracy at 99.23 %, recall value at 80.3 %, precision value at 68.78 %, and F1-score at 75.91 %. |
|---|---|
| ISSN: | 0045-7906 |
| DOI: | 10.1016/j.compeleceng.2025.110328 |