<![CDATA[Subjectivity in manual quality control for traditional roof tiles poses a significant challenge, as the current process relies on manual, visual inspection and subjective judgment. This research proposes an automatic system to classify tile quality from images using a Convolutional Neural Network (CNN), specifically the EfficientNet-B0 model enhanced with transfer learning. The study utilized a primary dataset comprising 616 local roof tile images collected directly from producers in Berjo Kidul, Godean, Yogyakarta. These images were manually labeled based on producer criteria into three distinct classes: 'Finished' (203 images), 'Underbaked' (213 images), and 'Broken/Cracked' (200 images). The methodology involved resizing all images to 224x224 pixels and applying data augmentation, including random rotation, horizontal flipping, and color jitter, to mitigate overfitting. The EfficientNet-B0 model, pre-trained on ImageNet, was implemented in PyTorch and trained for 10 epochs using an 80/20 train/validation split with the Adam optimizer. The model demonstrated outstanding performance, reaching 99.70% accuracy in validation. Further evaluation confirmed this robustness; the model perfectly identified the 'Underbaked' class and recorded only a single misclassification error on the test set. Qualitative analysis via a Flutter mobile application showed the system is resilient to changes in background and viewing angles, although its accuracy is compromised by poor lighting and strong shadows. This study validates the proposed system as a highly efficient and objective tool for a more reliable quality control process.]]>
