Automated Detection and Classification of Ceramic Surface Defects via a Hybrid Optical Imaging System and Optimized ResNet-50 Network
Duygu Demircan, Gulhan Ustabas Kaya, Rukiye Uzun
article (Uluslararası, Hakemli) | Yazar(lar): Duygu Demircan, Gulhan Ustabas Kaya, Rukiye Uzun | Yıl: 2026 | Dergi/Kurum: IEEE ACCESS | İndeks: WOS.SCI | DOI: 10.1109/access.2026.3713692 | Anahtar Kelimeler: Surfaces, Modeling, Ceramics, Training, Accuracy, Measurement, Defect detection, Inspection, Recording, Architec
The non-contact, rapid, and reliable detection of microscopic surface defects in ceramic materials prior to production remains a critical challenge in industrial optical inspection. This study proposes an innovative optical architecture, the Hybrid optical imaging system (HOIS), designed for high precision defect detection in ceramic components. The system integrates Lateral shearing digital holographic microscopy (LS-DHM) and Microscopic fringe projection profilometry (MFPP) within a unified configuration to simultaneously acquire phase and morphological surface information in a single measurement. The first objective is to retrieve phase distributions from holograms reconstructed by HOIS using the Hilbert Transform (HT) method, followed by the enhancement of defect contrast through the Sobel edge detection algorithm. The primary objective of the study is to use this data to perform a classification process with the deep learning-based to distinguish between defect-free and defective surfaces. In this context, the dataset was trained using the pre-trained ResNet-50 architecture through a two-stage evaluation process. The model trained exclusively with phase information extracted via HT achieved an accuracy of 85%, while the model incorporating edge-enhanced phase data reached an accuracy of 96%. The experimental findings confirm that the fusion of digital holographic imaging with appropriate edge preprocessing substantially improves deep learning-based defect classification. By integrating optical holography and artificial intelligence, the proposed HOIS provides a real-time compatible, non-contact, and high-accuracy inspection solution, contributing to enhanced industrial automation and optical quality assurance in ceramic manufacturing.