Hybrid Machine Learning for Automated Road Safety Inspection of Auckland Harbour Bridge DOI Open Access
Munish Rathee, Boris Bačić, Maryam Doborjeh

et al.

Electronics, Journal Year: 2024, Volume and Issue: 13(15), P. 3030 - 3030

Published: Aug. 1, 2024

The Auckland Harbour Bridge (AHB) utilises a movable concrete barrier (MCB) to regulate the uneven bidirectional flow of daily traffic. In addition risk human error during regular visual inspections, staff members inspecting MCB work in diverse weather and light conditions, exerting themselves ergonomically unhealthy inspection postures with added weight protection gear mitigate risks, e.g., flying debris. To augment inspections an using computer vision technology, this study introduces hybrid deep learning solution that combines kernel manipulation custom transfer strategies. video data recordings were captured conditions (under safety supervision industry experts) involving high-speed (120 fps) camera system attached vehicle. Before identifying hazard, unsafe position pin connecting two 750 kg segments MCB, multi-stage preprocessing spatiotemporal region interest (ROI) involves rolling window before frames containing diagnostic information. This ResNet-50 architecture, enhanced 3D convolutions, within STENet framework capture analyse data, facilitating real-time surveillance (AHB). Considering sparse nature anomalies, initial peer-reviewed binary classification results (82.6%) for safe (intervention-required) scenarios improved 93.6% by incorporating synthetic expert feedback, retraining model. adaptation allowed optimised detection false positives negatives. future, we aim extend anomaly methods various infrastructure enhancing urban resilience, transport efficiency safety.

Language: Английский

Detection of micro-pinhole defects on surface of metallized ceramic ring combining improved DETR network with morphological operations DOI Creative Commons

Y. Xiao,

Wang Xian, Yunlong Liu

et al.

PLoS ONE, Journal Year: 2025, Volume and Issue: 20(4), P. e0321849 - e0321849

Published: April 22, 2025

Metallized Ceramic Ring is a novel electronic apparatus widely applied in communication, new energy, aerospace and other fields. Due to its complicated technique, there would be inevitably various defects on surface; among which, the tiny pinhole with complex texture are most difficult detect, no reliable method of automatic detection. This Paper proposes detecting micro-pinhole surface metallized ceramic ring combining Improved Detection Transformer (DETR) Network morphological operations, utilizing two modules, namely, deep learning-based morphology-based defect detection detect pinholes, finally results such so as obtain more accurate result. In order improve performance DETR aforesaid module learning, EfficientNet-B2 used ResNet-50 standard network, parameter-free attention mechanism (SimAM) 3-D weight Sequeeze-and-Excitation (SE) linear combination loss function Smooth L1 Complete Intersection over Union (CIoU) regressive training network. The experiment indicates that recall precision proposed 83.5% 86.0% respectively, much better than current mainstream methods micro detection, meeting requirements at industrial site.

Language: Английский

Citations

0
DG2GAN: improving defect recognition performance with generated defect image sample DOI Creative Commons
Fuqin Deng,

Jialong Luo,

Lanhui Fu

et al.

Scientific Reports, Journal Year: 2024, Volume and Issue: 14(1)

Published: June 26, 2024

Abstract This article aims to improve the deep-learning-based surface defect recognition. In actual manufacturing processes, there are issues such as data imbalance, insufficient diversity, and poor quality of augmented in collected image for product A novel generation method with multiple loss functions, DG2GAN is presented this paper. employs cycle consistency generate images from a large number defect-free images, overcoming issue imbalanced original training data. DJS optimized discriminator introduced added encourage diverse images. Furthermore, maintain diversity generated while improving quality, new DG2 adversarial proposed aim generating high-quality The experiments demonstrated that produces higher greater compared other advanced methods. Using augment CrackForest MVTec datasets, recognition accuracy increased 86.9 94.6%, precision improved 59.8 80.2%. experimental results show using can obtain sample high employ augmentation significantly enhances technology.

Language: Английский

Citations

3
RADDA-Net: Residual attention-based dual discriminator adversarial network for surface defect detection DOI
Sukun Tian, Haifeng Ma, Pan Huang

et al.

Engineering Applications of Artificial Intelligence, Journal Year: 2024, Volume and Issue: 136, P. 108887 - 108887

Published: July 1, 2024

Language: Английский

Citations

3
Automated corner grading of trading cards: Defect identification and confidence calibration through deep learning DOI Creative Commons
Lutfun Nahar, Md. Saiful Islam, Mohammad Awrangjeb

et al.

Computers in Industry, Journal Year: 2024, Volume and Issue: 164, P. 104187 - 104187

Published: Sept. 19, 2024

Language: Английский

Citations

3
Lightweight defect detection network based on steel strip raw images DOI
Yue Huang, Zhen Chen, Zhaoxiang Chen

et al.

Engineering Applications of Artificial Intelligence, Journal Year: 2025, Volume and Issue: 145, P. 110179 - 110179

Published: Feb. 6, 2025

Language: Английский

Citations

0
LWMS-Net: A novel defect detection network based on multi-wavelet multi-scale for steel surface defects DOI
Xiaoyang Zheng, Weishuo Liu, Yan Huang

et al.

Measurement, Journal Year: 2025, Volume and Issue: unknown, P. 117393 - 117393

Published: March 1, 2025

Language: Английский

Citations

0
Broadband hybrid attention-based feature fusion network for printed circuit boards defect classification DOI
Feng Zhan, Weihan Qiu, Weiming Gan

et al.

Measurement, Journal Year: 2025, Volume and Issue: unknown, P. 117363 - 117363

Published: April 1, 2025

Language: Английский

Citations

0
An adversarial network based on anomaly domain decomposition and transformation for industrial PCBA defect inspection DOI

Qiuling Pan,

Teng Liu, Yue Hou

et al.

Neural Computing and Applications, Journal Year: 2025, Volume and Issue: unknown

Published: April 9, 2025

Language: Английский

Citations

0
Eternal-MAML: a meta-learning framework for cross-domain defect recognition DOI Creative Commons
J. H. Feng, Haigang Zhang, Zhifeng Wang

et al.

PeerJ Computer Science, Journal Year: 2025, Volume and Issue: 11, P. e2757 - e2757

Published: May 7, 2025

Defect recognition tasks for industrial product suffer from a serious lack of samples, greatly limiting the generalizability deep learning models. Addressing imbalance defective samples often involves leveraging pre-trained models transfer learning. However, when these models, on natural image datasets, are transferred to pixel-level defect tasks, they frequently overfitting due data scarcity. Furthermore, significant variations in morphology, texture, and underlying causes defects across different products lead degradation performance, or even complete failure, directly transferring classification model trained one type another. The Model-Agnostic Meta-Learning (MAML) framework can learn general representation multiple build foundational model. Despite lacking sufficient training data, MAML still achieve effective knowledge among cross-domain tasks. We noticed there exists label arrangement issues because random selection which seriously affects performance during both testing phase. This article proposes novel framework, termed as Eternal-MAML, guides update classifier module by meta-vector that shares commonality batch inner loop, addresses phenomenon caused phase vanilla MAML. Additionally, feature extractor this combines advantages Squeeze-and-Excitation Residual block enhance stability improve generalization accuracy with learned initialization parameters. In simulation experiments, several datasets applied verified meta-learning proposed Eternal-MAML framework. experimental results show outperforms state-of-the-art baselines terms average normalized accuracy. Finally, ablation studies conducted examine how primary components affect its overall performance. Code is available at https://github.com/zhg-SZPT/Eternal-MAML .

Language: Английский

Citations

0
An Economic Evaluation for Implementation of Zero Defects and Zero Waste Inspection Solution in the Wind Energy Manufacturing Industry DOI
Joan Lario, Julián Luengo, Raúl Poler

et al.

Lecture notes on data engineering and communications technologies, Journal Year: 2025, Volume and Issue: unknown, P. 245 - 251

Published: Jan. 1, 2025

Language: Английский

Citations

0