Classificação de telhas cerâmicas tipo romana no processo de fabricação industrial com redes neurais convolucionais

Abstract

The incorporation of emerging technologies, such as Artificial Intelligence (AI), into production processes has driven significant advances in efficiency, automation, and quality control across various industrial sectors. In the context of the ceramic industry, product quality is an essential requirement to ensure structural integrity, durability, and competitiveness in the construction sector. However, quality control is still predominantly performed through manual visual inspection, which is subject to operator subjectivity and fatigue, compromising the reliability of the results. This work aimed to develop and validate a Convolutional Neural Network (CNN) model for the automatic classification of visual defects in Roman-type ceramic roof tiles, using images captured directly from the production line, in alignment with the principles of Industry 4.0. To this end, a balanced dataset of 1,600 images was built, consisting of half normal tiles and half defective ones, which underwent a 12-step preprocessing pipeline, including background removal, histogram equalization, pixel normalization, and data augmentation using the Albumentations library. The CNN architecture was composed of 12 layers, including convolutional, pooling, and dense layers, with a final sigmoid activation function. The performance was evaluated using metrics such as accuracy, precision, recall, F1-score, and confusion matrix. The model achieved an accuracy of 99.69% and an F1-score of 0.997, recording only one false negative and no false positives, outperforming related works in the literature and demonstrating high robustness in the face of morphological and uncontrolled lighting variations. It is concluded that the proposed approach is effective in replacing subjective manual inspections with an automated system, providing greater efficiency, traceability, and a reduction in human errors.