| Creators: |
Kostarev, Pavel and Bartel, Alexander |
| Title: |
A Modular Multimodal Pipeline for Scratch Detection, Segmentation, and Coordinate Extraction on Polymer Ski Surfaces |
| Item Type: |
Conference or Workshop Item |
| Event Title: |
12th International Conference on Control, Decision and Information Technologies (CoDIT) |
| Event Location: |
Bari, Italy |
| Event Dates: |
13-16 July 2026 |
| Page Range: |
pp. 548-553 |
| Additional Information: |
10.1109/CoDIT70676.2026.11630706 |
| Date: |
7 August 2026 |
| Divisions: |
Informationsmanagement |
| Abstract (ENG): |
We present a system that can automatically locate and extract scratches from plastic ski surfaces using artificial intelligence (AI). Using a custom dataset consisting of 180 pairs of 2D grayscale and 3D depth profile images, we designed and compared three scratch detection and segmentation pipelines. The resulting segmented defects were mapped to X and Y coordinates for use in subsequent production processes. The most effective strategy fused computer vision (CV) algorithms with a multimodal machine learning architecture comprising a You Only Look Once (YOLO) module for scratch detection, a UNet++ module for scratch segmentation, and a region-growing algorithm to enhance segmentation. Our best approach achieved mean Average Precision (mAP) of 91% computed at Intersection over Union (IoU) threshold of 0.5 over all defect classes when running on the test set. This approach outperformed the other two in terms of time, performance, and accuracy, while also requiring minimal hardware. The proposed methodology is also parameterised to provide flexibility and allow adjustments for various industry applications. The modular nature of the approach allows new algorithms to be integrated and further pre- and post-processing to be carried out. |
| Forthcoming: |
No |
| Main areas or research: |
Transformationmanagement |
| Language: |
English |
| Uncontrolled Keywords: |
Anomaly detection, Scratch detection, Segmentation models |
| Link eMedia: |
Download |
| Citation: |
Kostarev, Pavel and Bartel, Alexander
(2026)
A Modular Multimodal Pipeline for Scratch Detection, Segmentation, and Coordinate Extraction on Polymer Ski Surfaces.
In: 12th International Conference on Control, Decision and Information Technologies (CoDIT), 13-16 July 2026, Bari, Italy, pp. 548-553.
ISBN 979-8-3195-2077-7
|
 |
View Item in edit mode (academic staff only) |
