LEE- Tekstil Mühendisliği Lisansüstü Programı

Bu topluluk için Kalıcı Uri

http://hdl.handle.net/11527/19465

Gözat

Yazar "Birsen, Sinem" ile LEE- Tekstil Mühendisliği Lisansüstü Programı'a göz atma
  • Öge
    Detection and classification of fabric defects with an innovative model and perspective
    (Graduate School, 2025-01-24) Birsen, Sinem ; Sarıçam, Canan ; 503211815 ; Textile Engineering
    The global textile and ready-wear industries represent a substantial and highly competitive sector of the global market, regulated by both quality and price. Fabric defects significantly impact the quality of ready-wear items. Likewise, undetected defects during traditional quality control processes can lower the value of fabrics. Consequently, detecting and classifying fabric defects becomes crucial for companies aiming to remain competitive in the market, whether through quality leadership or price rivalry. While the literature includes various studies on fabric defect detection and classification, these often rely on open-source or custom datasets and employ well-known deep learning architectures or propose novel architectures. However, no study to date has specifically accounted for the structural differences between woven and knitted fabrics when designing models for fabric defect detection and classification. To address this gap, the present study developed two new datasets – the Woven Fabric dataset and the Knitted Fabric dataset – and designed a novel deep learning architecture using Convolutional Neural Networks (CNNs). As the study method, an open-source dataset (TILDA) was first utilized to evaluate well-known architectures (VGG19, ResNet50, InceptionV3) and inspire the design of a custom CNN model. This custom architecture was then optimized using 3-factor, 2-level factorial design experiments to refine structural parameters. The model's performance was validated on three custom datasets (Woven, Knitted, and Woven-Knitted Fabric datasets). Subsequently, the hyperparameters affecting model performance were optimized using a 4-factor, 2-level factorial design, and the model was revalidated on both open-source and custom datasets. The model was evaluated using additional metrics, including recall, precision, specificity, and F1-score, demonstrating superior performance. Training performance was analyzed using Accuracy/Loss curves, confirming no signs of overfitting. Furthermore, confusion matrixes indicated the model's effectiveness and robustness in classifying different defect classes. The final model achieved 97.37% accuracy on the TILDA dataset, 97.73% accuracy on the Woven Fabric dataset, 96.92% accuracy on the Knitted Fabric dataset, and 98.36% accuracy on the Woven-Knitted Fabric dataset. The results including recall, precision, specificity, and F1-score all suppressed the expected criteria. These results demonstrate the model's capability to detect and classify fabric defects effectively, accounting for the structural differences between fabric types. Future studies will focus on expanding the datasets to include more fabric types and design samples, aiming to enhance the model's generalization ability and performance across different fabric domains. In conclusion, this study successfully identified a custom CNN model suitable for both woven and knitted fabric types, considering their structural differences. It lays the foundation for future research and industrial implementation in automated fabric quality control.