Akademik Veri Yönetim Sistemi
IEEE Access, cilt.14, ss.18272-18284, 2026 (SCI-Expanded, Scopus)
Automating defect classification in electrospun polyacrylonitrile (PAN) nanofibers is essential for improving quality control and reducing the experimental effort required for scanning electron microscopy (SEM) based inspection. This study introduces a domain specific deep learning model for the accurate classification of nanofiber defects using a uniquely curated SEM dataset labeled into slightly defective, defective, and non-defective categories. Unlike conventional binary approaches, the proposed framework incorporates an intermediate class that captures subtle morphological deviations that are important for early stage defect recognition. A custom Nanofiber Classifier is developed and enhanced with a multiscale Inception module together with hyperparameters optimised through a genetic algorithm (GA), enabling expressive and discriminative feature extraction across heterogeneous fiber geometries. Evaluation against widely used pretrained architectures such as GoogleNet and AlexNet shows that the proposed model, although developed and trained entirely from scratch, achieves exceptional classwise performance with accuracy values of 98.64% for slightly defective, 99.16% for defective, and 99.18% for non defective samples. These findings demonstrate that a purpose built multiscale architecture provides an effective, reliable, and scalable solution for SEM based nanofiber quality assessment.