Detection of brick firing defects using image processing technology and deep learning

With the decline in the working-age population, automation of production processes has become increasingly urgent in the secondary industry. While automation of visual inspection has been widely studied, most research relies on public datasets; few studies use data acquired on real production lines. Therefore, in this study, with the cooperation of a company that produces refractory bricks, we developed an anomaly detection system by acquiring RGB images of bricks on an actual production line from multiple angles and with multiple lighting patterns. The proposed system first performs “primary detection” using traditional image processing techniques (binarization, line detection, etc.) to extract potential defects. It then carries out “secondary classification” via deep learning to confirm the presence or absence of defects. This two-step approach achieves high detection accuracy while strictly avoiding missed defects. Experimental results showed that Accuracy 0.983, Recall 1.0, Precision 0.860, and F value 0.924 were achieved, indicating that it is possible to construct a system that can withstand actual operation.