Abstract:
The global market for ceramic tiles industry is highly competitive nowadays. Quality
control in the production process in the ceramic tile industry has been a key factor for
retaining existence in such a competitive market. Deep learning-based ceramic tile
inspection systems are very useful in this respect because the manual inspection is time consuming and not accurate enough. Hence, deep learning can help ceramic tile inspection
system faster and accurate. Two difficult problems are mainly posed by deep learning based ceramic tile inspection systems. They are defect detection and defect detection
classification. Even though there has been plenty of research addressing the defect
detection problem, the research aiming at solving the classification problem is scarce.
Moreover, in this research, we used two models to compare with our proposed variation
CNN model to find out which is the better one to identify defect detection. We have found
four types of defected tiles including multi-label defected tiles. In this research, first we
used VGG-16 for training and we got 54% accuracy which was not good enough. Then we
tried another model for training, Inception-v3 gave us an optimistic result on training
dataset which was 95%. Then we have used our proposed CNN model on the tiles dataset
and we got 96% accuracy for training datasets of images. Even though Inceptionv3 has
better accuracy on training datasets but for testing datasets, it gives a poor result of 33%,
On the other hand with our proposed variation of CNN model we can identify defected
tiles by 91%. Finally, this thesis paper focuses and proposes technical aspects of tiles defect
detection in a faster and easier way by comparing our proposed variation of the CNN model
with other pre-trained models.