Classification of biscuit tiles for defect detection using Fourier transform features

Automated defect detection is difficult to achieve in ceramic tile manufacturing today. Computer vision and machine learning based approaches are commonly utilised for this purpose. This paper considers the problem of defect detection in the textured ceramic tiles quality analysis. Instead of detecting defects on the finished tile, the biscuit tile is considered, a pressed, dried, decorated tile before its firing in the kiln. As it is an intermediary product during tile production, classifying them as defected or not before the firing can significantly reduce energy and material costs. To this end, in this paper we propose a new Fourier spectrum annuli feature extraction method. It is based on Fourier spectrum of the surface biscuit tile image and tested on real tile examples from the ceramic tile industry. According to the observed results, it outperforms several well-known methods for feature extraction on real-world tile datasets reaching an F1 score of 0.9236 and 0.8866 on the Black Random Stripes and Stripes Brown Light tile designs respectively.