<p>Steel skull formation on the refractory walls of steelmaking equipment is a critical issue that affects both process efficiency and product quality. In the RH (Ruhrstahl Heraeus) refining process, such steel skulls can alter molten steel composition, increase non-metallic inclusions, and raise deoxidizer demand. This study aimed to quantitatively evaluate steel skull formation and analyze its causal relationship with process variables during the RH refining process. Internal images of an RH vessel were collected using a compact vision camera mounted on the top-oxygen blowing lance. The images were then analyzed using a deep learning-based object detection technique to quantitatively measure the cross-sectional area of steel skull within the RH vessel. To capture the complex nonlinear characteristics inherent in the RH process, a combined model of the Generalized Propensity Score and the Generalized Additive Model was constructed. This approach identified that skull formation is influenced by ferro-alloy input and RH arrival temperature, with distinct effects observed depending on the deoxidation method and process conditions. The results of this study are expected to contribute to process optimization aimed at minimizing steel skull formation during RH operations, while improving the quality of the molten steel produced.</p>