Institutional Repository of Vilnius University
Defektų aptikimas medžiagų paviršiuose /
Abstract
dc:descriptionThe focus of this study is on material surface defect segmentation, with the aim of comparing the effectiveness of different models developed and adapted for defect segmentation of steel surfaces. The models used for this comparative analysis are different convolutional neural network architectures and a visual transformer architecture, and their performance is compared with the state-of-the-art model that achieved an accuracy of 98.74% . By evaluating the effectiveness of these models, this work identifies the most suitable models for surface defect detection and segmentation and provides insights for improving the accuracy and efficiency of future surface defect detection systems. The purpose is that. Quality checks are essential in any manufacturing process and ensuring the quality of the final product is essential to maintain customer satisfaction and brand reputation. However, quality checks are time consuming, labour intensive and subject to human error. Integrating artificial intelligence into quality checks improves accuracy, speed and consistency. AI-powered quality inspections can detect defects that human inspectors might miss, ultimately improving product quality and reducing costs associated with defects and recalls. Changing the weight initialisation from Xavier to Kaiming, applying dataset normalisation and dropout regularisation to the SegDecNet architecture can improve the performance of this architecture. The ViT architecture models can achieve better results than the CNN architecture models. The best combination of encoder and decoder is a PAN encoder with a mobileVitv2 decoder. With this combination, better results can be achieved for the semantic segmentation of steel defects task.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Šaltenis, Bartas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD192846203&prefLang=en_US