N.A. Andriyanov1, A.A. Volnenko2, V.E. Dementiev3

1,2 Financial University under the Government of the Russian Federation (Moscow, Russia)

3 Ulyanovsk State Technical University (Ulyanovsk, Russia)

Problem setting. At present, the monitoring of various objects using non-invasive, that is, without direct contact with the object, methods is being actively developed. However, such an approach, firstly, requires high-quality equipment for photo and video recording, and, secondly, imposes various high requirements on the algorithms that implement such monitoring. In particular, there should be a sufficiently high accuracy of defect recognition in images of iron products. However, on the basis of such an information-measuring system, it will also be possible to offer control decisions for the replacement of one or another structural element, for example, or to detect defects in production.

Target. The main goal of this work is to improve the quality of defect recognition in images of metal products through the use of computer vision transformer architectures. From the presented goal, such tasks arise as searching for a dataset for training models, marking it up, training a transformer architecture and comparing it with other approaches.

Results. Algorithms for segmenting steel images based on convolutional neural networks have been developed that provide the Dice-score metric at the level of 70%. It was proposed to use the SegFormer architecture to improve the quality of segmentation. A new image processing algorithm based on the SegFormer architecture has been developed, which made it possible to increase the Dice-score metric to 74%.

Practical significance. The developed algorithms for recognizing defects in steel products can be used in a number of applied problems of image analysis, such as monitoring the state of metal structures, decision support information systems for replacing structural elements, identifying defective steel products, and others.

Andriyanov N.A., Volnenko A.A., Dementiev V.E. Monitoring of the condition of metal products based on computer vision systems. Neurocomputers. 2023. V. 25. № 5. Р. 50-57. DOI: https://doi.org/10.18127/j19998554-202305-07 (In Russian)

1. Fedosov A.V., Gainullina L.A. Methods of non-destructive testing. Electrical and information complexes and systems. 2015. V. 11. № 2. P. 73–78. (in Russian)
2. Vasiliev A.A., Gorin L.N., Igoshin D.N. Non-destructive testing and the possibility of its application in car repair production. Academy. 2017. V. 1. № 6(21). P. 32–33. (in Russian)
3. Lee S., Chung Y., Kim W. Defect Recognition and Morphology Operation in Binary Images Using Line-Scanning-Based Induction Thermography. Applied Sciences. 2022. V. 12. P. 6006. DOI 10.3390/app12126006.
4. Kou J., Wang J., Ding J., Ge X. Spatial Simulation and Prediction of Land Use/Land Cover in the Transnational Ili-Balkhash Basin. Remote Sensing. 2023. V. 15. P. 3059. DOI 10.3390/rs15123059.
5. Zhang Y., Dian Y., Zhou J., Peng S., Hu Y., Hu L., Han Z., Fang X., Cui H. Characterizing Spatial Patterns of Pine Wood Nematode Outbreaks in Subtropical Zone in China. Remote Sensing. 2021. V. 13. P. 4682. DOI 10.3390/rs13224682.
6. Andriyanov N.A., Vasiliev K.K. Use autoregressions with multiple roots of the characteristic equations to image representation and filtering. Proceedings of the International Conference on Information Technology and Nanotechnology. 2018. P. 273–281. DOI 10.18287/1613-0073-2018-2210-273-281.
7. Vasiliev K.K., Dementiev V.E., Andrianov N.A. Analysis of the effectiveness of estimating the changing parameters of a doubly stochastic model. Radio Engineering. 2015. № 6. P. 12–15. (in Russian)
8. Vasiliev K.K., Dementiev V.E., Andrianov N.A. Detection of extended signals against the background of doubly stochastic images. Radio Engineering. 2016. № 9. P. 23–27. (in Russian)
9. Andreeva O.V., Tsilikh A.S. Recognition of defects in steel using neural networks and problems of their classification. Collection of materials XXVI International Scientific-technical. conf. "Information systems and technologies IST–2020". Nizhny Novgorod: Nizhny Novgorod State Technical University named after R.E. Alekseev. 2020. P. 919–923. (in Russian)
10. Gaidar A.I., Yakimov P.Yu., Viktorenkov A.E., Shustanov A.V. Detection and recognition of defects of the inner surface of metal pipes. Proceedings on the materials of the VI International Conference and youth school "Information technologies and nanotechnology" (ITNT–2020). Samara: Samara National Research University named after Academician S.P. Korolev. 2020. V. 4. P. 741–749. (in Russian)
11. Girshick R. Fast R-CNN. IEEE International Conference on Computer Vision. Santiago, Chile. 2015. P. 1440–1448. DOI 10.1109/ICCV.2015.169.
12. Andriyanov N., Dementiev V., Tashlinsky A. Deep Markov Models of Multidimensional Random Fields. Procedia Computer Science. 2020. V. 176. P. 1289–1298. DOI 10.1016/j.procs.2020.09.138.
13. Ronneberger O., Fischer P., Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention. 2015. V. 9351. P. 1202–1216. DOI 10.1007/978-3-319-24574-4_28.
14. Xie E., Wang W., Yu Z., Anandkumar A., Alvarez J.M., Luo P. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. ArXiv. 2021. V. abs/2105.15203.
15. Andriyanov, N.A., Dementiev V.E., Tashlinsky A.G. Object detection in an image: from Bayes and Neumann–Pearson criteria to detectors based on EfficientDet neural networks. Computer Optics. 2022. V. 46. № 1. P. 139–159. DOI 10.18287/2412-6179-CO-922. (in Russian)
16. Severstal: Steel Defect Detection | Kaggle. [Electronic resource] – Access mode: https://www.kaggle.com/c/severstal-steel-defect-detection, date of reference: 16.06.2023.