A.O. Burygin1, I.G. Panin2

1 CJSC "NORSI-TRANS" (Moscow, Russia)

2 Kostroma State University (Kostroma, Russia)

1 g.t.m.p@yandex.ru, 2 igpanin@list.ru

The main problem of applying neural networks in the task of defect detection is that there is a limitation in the size of the set of annotated defect images compared to the set of defect-free images. This imbalance in the dataset poses a challenge for the model to effectively learn and generalize patterns associated with defects. The scarcity of annotated defect samples hinders the model's ability to capture the diverse characteristics of anomalies, making it prone to misdetection and reduced performance on real-world data. To solve this problem, the authors have proposed a neural network architecture based on two neural networks U-Net++ and PatchCore trained in the student-teacher paradigm. In this paradigm, the student model undergoes initial training exclusively on defect-free images. Consequently, disparities emerge in the learned features between the student and teacher models, facilitating the computation of an anomaly metric and streamlining the anomaly localization process. The aim of the research is to improve the efficiency of defect detection by creating methods and algorithms, and dataset augmentation systems, as an option to solve the problem of defect-free images. The overarching goal of this research is to enhance the efficacy of defect detection through the development of methodologies, algorithms, and dataset augmentation systems. This serves as a strategic approach to address the inherent challenge posed by the scarcity of annotated defect images in comparison to their defect-free counterparts. By creating methods and augmenting datasets, we aim to mitigate the limitations imposed by the uneven distribution of defect and defect-free samples, ultimately advancing the robustness and performance of defect detection models. A dual system of defect detection based on the U-Net++ pre-trained network and feature estimation of individual fragments based on the PatchCore network is presented. The research incorporates computational experiments designed to assess the effectiveness of the proposed method in comparison to existing counterparts. An in-depth efficiency analysis has been conducted, focusing on dataset enrichment using the generative-adversarial network ResStyleGAN. This examination aims to provide empirical evidence of the method's performance and its potential advantages over established approaches in the context of defect detection. The method developed, centered around the dual system PatchNet++ and preliminary generation of new defects using ResStyleGAN, is positioned for practical application in real production scenarios. Subsequent stages involve a detailed analysis to evaluate its performance and suitability in real-world settings.

Burygin A.O., Panin I.G. Dual defect detection and generation system in planar surface. Neurocomputers. 2024. V. 26. № 3. Р. 55-67. DOI: https://doi.org/10.18127/j19998554-202403-06 (In Russian)

1. Vlasov A.I., Zavyalov N.V., Selivanov K.V., Skalchenkov I.I. Application of neural networks in detecting PCB defects. Neurocomputers. 2022. V. 24. № 6. Р. 5–19. DOI 10.18127/j19998554-202206-01. (In Russian)
2. Burygin A.O., Panin I.G. Detection of sling defects based on unmarked data. Collection of articles XI International Scientific and Practical Conference "Actual scientific research". Penza: Science and Education. 2023. P. 94–97. (In Russian)
3. Bukalov G.K., Burygin A.O., Panin I.G., Tortsev A.V. Segmentation and Defect Detection on Large Areas in Several Frames. Information Technologies. 2022. V. 28. № 1. P. 20–25. DOI 10.17587/it.28.20-25. (In Russian)
4. Bergmann P., Fauser M., Sattlegger D, Steger C. Uninformed Students: Student–Teacher Anomaly Detection with Discriminative Latent Embeddings. [Electronic resource] – Access mode: https://arxiv.org/pdf/1911.02357, date of reference 13.02.2024.
5. Deng H., Li X. Anomaly Detection via Reverse Distillation from One-Class Embedding. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA. 2022. P. 9727–9736. DOI 10.1109/CVPR52688.2022.00951.
6. Schlegl T., Seeböck P., Waldstein S.M., Schmidt-Erfurth U., Langs G. Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery. International Conference on Information Processing in Medical Imaging. 2017. V. 10265. P. 146–157. DOI 10.1007/978-3-319-59050-9_12.
7. Roth K., Pemula L., Zepeda J., Schölkopf B., Brox T., Gehler P. Towards Total Recall in Industrial Anomaly Detection. Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA. 2022. P. 14298–14308. DOI 10.1109/CVPR52688.2022.01392.
8. Lee S., Lee S., Song B.C. CFA: Coupled-Hypersphere-Based Feature Adaptation for Target-Oriented Anomaly Localization. IEEE Access. 2022. V. 10. P. 78446–78454. DOI 10.1109/ACCESS.2022.3193699.
9. Deng J., Guo J., Yang J., Xue N., Kotsia I., Zafeiriou S. ArcFace: Additive Angular Margin Loss for Deep Face Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2022. V. 44. № 10. P. 5962–5979. DOI 10.1109/tpami.2021.3087709.
10. Efe U., Ince K.G., Alatan A.A. DFM: A Performance Baseline for Deep Feature Matching. Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Nashville, USA. 2021. P. 4279–4288. DOI 10.1109/CVPRW53098.2021.00484.
11. Batzner K., Heckler L., König R. EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies. Winter Conference on Applications of Computer Vision (WACV). Waikoloa, USA. 2024. P. 127–137. DOI 10.1109/WACV57701.2024.00020.
12. Burygin A.O., Panin I.G. Defect image generation using a variation of generative-adversarial network based on stylegan3. Collection of articles of the IV International Scientific and Practical Conference "Fundamental and Applied science: topical issues of theory and practice". Penza: Science and Education. 2023. P. 57-64. (In Russian)
13. Zhu T., Chen J., Zhu R., Gupta G. StyleGAN3: Generative Networks for Improving the Equivariance of Translation and Rotation. [Electronic resource] – Access mode: https://arxiv.org/pdf/2307.03898v1, date of reference 23.03.2024.