Due to the high hardness and brittleness of ceramic materials, conventional cutting methods result in poor quality and machining difficulties. Additive manufacturing has also been tried in various ways, but it has many limitations. This study aims to propose a system to monitor surface defects that occur during the printing process based on high-viscosity composite resin that maximizes ceramic powder content in real time using image processing and convolutional neural network (CNN) algorithms. To do so, defects mainly observed on the surface were classified into four types by form: pore, minor, critical, and error, and the effect of each defect on the printed structure was tested. In order to improve the classification efficiency and accuracy of normal and defective states, preprocessing of images obtained based on cropping, dimensionality reduction, and RGB pixel standardization was performed. After training and testing the preprocessed images based on the DenseNet algorithm, a high classification accuracy of 98% was obtained. Additionally, for pore and minor defects, experiments confirmed that the defect surfaces can be improved through the reblading process. Therefore, this study presented a defect detection system as well as a feedback system for process modifications based on classified defects.
Keywords: CNN; additive manufacturing; ceramic 3D printing; composite resin; high-viscosity resin; photo-polymerization; surface defects.