Study on the Correlation between As-Cast Surface Roughness and Pore Size Using Trained Semantic Segmentation Model

As-cast surface roughness is a crucial quality indicator for cast parts, representing not only the surface quality of the product but also serving as an important marker for analyzing the internal microstructure of the casting. Consequently, an inline measurement method for as-cast surfaces is essential. However, traditional methods such as visual inspection and stylus profilometry face challenges in automation, while advanced laboratory methods like Laser Scanning Confocal Microscopy (LSCM) and computer vision systems with optical images encounter limitations in measurement speed or accuracy. In response to these challenges, we propose a training strategy based on observations: the pores in the web-like topography of the as-cast surface tend to enlarge as the surface roughness increases. To leverage this observation, we selected and trained two segmentation models, Unet and DeepLabv3+, to identify pores on the as-cast topography images and to calculate their sizes, thereby predicting the corresponding as-cast surface roughness. Our modified Unet model demonstrated superior performance in pore segmentation, enabling us to calculate pore area and diameter accurately. A high Pearson correlation coefficient of0.956 between the average pore diameter and the values of as-cast surface roughness $S_{a}$, arithmetical mean height, confirmed our hypothesis. This segmentation technique not only predicts the as-cast surface roughness but also holds potential for identifying other casting surface defects, such as cold shut and porosity.