ISCSR Research Publishing
Frontiers of Innovation in Science
Research Article Published October 9, 2026 Article fis20260009

Defect-Preserving Token Allocation for Industrial Surface Inspection

Read & download PDF
Abstract

Industrial surface inspection requires fast and reliable recognition of small defects such as scratches, pits, stains, cracks, and coating bubbles. Vision Transformer models can capture long-range visual context, but their full-token computation is often unnecessary for large uniform regions on steel plates, fabrics, printed circuit boards, and coated surfaces. This study investigates defect-preserving token allocation for efficient industrial inspection. We propose the Defect Token Allocation Network (DTAN), which assigns computational priority according to local anomaly contrast, patch-level uncertainty, and cross-layer defect persistence. Instead of using a fixed pruning ratio, DTAN keeps suspected defect patches at full attention depth, compresses repetitive background patches, and rechecks borderline patches through a lightweight verification branch. The study used inspection data collected from 14 production lines, including 286,000 high-resolution images and 412,600 annotated defect instances. Defect sizes ranged from 6 × 8 pixels to 420 × 360 pixels, which allowed evaluation on both tiny and large surface anomalies. On the NEU-Surface, DAGM, KolektorSDD2, and internal factory datasets, DTAN reached 91.8% defect-level F1-score while reducing transformer computation by 33.9%. For defects smaller than 32 × 32 pixels, recall remained at 88.6%, compared with 82.4% for a standard token-pruning model. In online testing, average inspection latency decreased from 47.3 ms to 29.5 ms per image, allowing stable processing at 1.6 m/s conveyor speed. False alarms caused by repetitive texture were reduced by 21.7%. These results show that defect-aware token allocation can improve the deployment efficiency of transformer inspection models without losing small industrial defects.

Keywords
Industrial inspectionsurface defect detectionVision Transformertoken allocationefficient inferenceanomaly localizationquality control
References
  1. Xiong, W., Guo, Y., & Zeng, Y. (2026). A Study on the Energy Efficiency of MEP Systems and Coordinated Dispatch Mechanisms for Multi-energy Systems in High-density Urban Buildings. Available at SSRN 7120518.
  2. Olianin, D., Tsupryk, H., Hovorushchenko, T., Bahrii-Zaiats, O., & Andrushchak, I. (2025, June). Transformer neural networks in Industry 4.0. In CITI (pp. 331-336).
  3. Hong, Z., Liang, S., Xu, T., & Chen, H. (2026). A Study on Failover Verification and Recovery Objective Prediction for Cross-Region Cloud Services.
  4. Zhang, Z., Wang, J., Li, Z., Wang, Y., & Zheng, J. (2025). Anncoder: A mti-agent-based code generation and optimization model. Symmetry, 17(7), 1087.
  5. Maiga, O., & Akhloufi, M. A. (2026, July). Unsupervised Rice Leaf Disease Detection Using PatchCore Feature Memory Banks. In 2026 IEEE International Conference on Signals and Systems (ICSigSys) (pp. 1-5). IEEE.
  6. Huang, J., Yin, J., Xu, T., & Yang, J. (2026). Quality Assessment and User Adoption Prediction for Enterprise-Level Financial Reporting Data Products. Available at SSRN 7179118.
  7. Yao, G., Zheng, J., Wang, Z., Zhang, W., Han, R., Zhao, C., ... & Liu, R. (2026, March). V-pruner: A fast and globally-informed token pruning framework for vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 40, No. 40, pp. 34396-34404).
  8. Chen, F., Li, S., Liang, H., Xu, P., & Yue, L. (2025). Optimization Study of Thermal Management of Domestic SiC Power Semiconductor Based on Improved Genetic Algorithm.
  9. Barusco, M., Pezze, D. D., Borsatti, F., & Susto, G. A. (2026). AdapTS: Lightweight Teacher-Student Approach for Multi-Class and Continual Visual Anomaly Detection. arXiv preprint arXiv:2603.17530.
  10. Zhang, Z. (2026). A Study on the Impact of Cost Presentation on Decision-making Bias and Cancellation Behavior in Online Subscriptions. Available at SSRN 7349839.
  11. Qi, C., & Qiao, X. (2026). Building and Operating a Large Scale Multi-Agent System: A Case Study from Industry. Available at SSRN 6795198.
  12. Du, Y., Liu, Q., Dong, Y., & Chen, Z. (2026). Cross-Domain Transfer and Few-Shot Recognition of Defect Images in Complex Industrial Settings.
  13. Ercoşkun, B. (2025). Integration of Image Processing Techniques into Automotive Quality Control Processes: Methods, Applications, and Future Perspectives. Gazi Journal of Engineering Sciences, 11(3), 426-448.
  14. Zhao, J., Fan, J., & Li, L. (2026). A Study on an Explainable Causal-Enhanced LLM Agent for Predicting the Forming Quality of Automotive Component Materials.
  15. You, S. (2026). Verifiable Audit Mechanisms in AI Compliance Automation: Scalability. Available at SSRN 6547458.
  16. Bao, Y., & Wang, H. (2026). Orion: A High-Throughput, Fault-Tolerant Event Routing Architecture for Hyperscale Data Streams. Fault-Tolerant Event Routing Architecture for Hyperscale Data Streams (January 01, 2026).
  17. Cahyani, P. I. G., Suartana, K. D. D., & Yudistira, N. (2025). Input-adaptive visual preprocessing for efficient fast vision-language model inference. arXiv preprint arXiv:2512.20839.
  18. Wang, S., Feng, Y., & Fang, X. (2026, May). A Large Language Model-Enabled Multi-Agent Collaboration Method for Complex Task Solving. In 2026 6th International Symposium on Computer Technology and Information Science (ISCTIS) (pp. 253-256). IEEE.
  19. Yang, Q., & Du, Y. (2026). Predicting Rollback Risks and Controlling Gradual Rollout Traffic for Online Ranking Feature Deployments.
  20. Pradeep, A., Nazari, S., Taheri, M., & Herglotz, C. (2026). Fusion: A Framework for Unified Sequential Token AdaptatIon in VisiOn TraNsformers. arXiv preprint arXiv:2607.02612.
  21. Zhao, Z., & Welsch, R. E. (2026). Point-in-Time Financial RAG with Frozen LLMs and Market-Feedback Adaptive Retrieval. arXiv preprint arXiv:2605.31201.
  22. Yuan, Y., Huang, J., Yin, J., & Xu, T. (2026). Confidence Calibration and Semantic Error Analysis for Ambiguous Coreference Resolution in User-generated Social Media Text. Available at SSRN 7393238.
  23. Latibari, B. S., Homayoun, H., & Sasan, A. (2025). Optimizing vision transformers: Unveiling ‘focus and forget’for enhanced computational efficiency. IEEE Access, 13, 27908-27927.
  24. Jiao, Y., Shi, T., Zhao, B., & Wang, A. (2026). Retrieval-Guided Structured Reasoning and Interpretable Representation Learning for Large-Scale Video–Language Models.
  25. Liang, S., Du, Y., Chen, W., & Liu, Z. (2026). Order Allocation and Emergency Transportation Decisions Amid Multi-Source Procurement Disruptions.
  26. Islam, M. S., Bappy, M. M., Tushar, S. R., Zamiela, C., & Wasi, A. T. (2026). Beyond local receptive fields: vision transformers for real-time surface defect detection in FDM. The International Journal of Advanced Manufacturing Technology, 143(9), 5399-5420.
  27. Liang, S., Hsu, K. H., & Liu, Z. (2026). Rolling Scheduling and Capacity Balancing for Server Assembly Under Engineering Change Disturbances.
  28. Bao, Y., & Wang, H. (2026). Large-scale Metrics Migration: A Systematic Approach to Rebuilding Production Measurement Infrastructure. Available at SSRN 7185660.
  29. Gnieciak, D., & Szandala, T. (2025). Large language models versus static code analysis tools: A systematic benchmark for vulnerability detection. IEEE Access, 13, 198410-198422.
  30. Yuan, Y., Yin, J., Huang, J., & Xu, T. (2026). Weakly Supervised Anomaly Detection and Privacy Risk Scoring in Community Message Threads.
Publication details
Journal
Frontiers of Innovation in Science
Volume
1 (2026)
Issue
1 · Forthcoming issue
Article number
fis20260009
License
CC BY 4.0