Abstract: Accurate identification of complex multiphase inclusions is critical for the development of environmentally friendly Bi-based free-cutting steels. Traditional metallographic microscopy methods fail to reliably classify and quantify Bi-alloyed free-cutting phases, while existing commercial automated inclusion analysis systems still show limited capability in identifying MnS-Bi composite inclusions with complex morphologies and coexisting bright/dark contrasts. Based on the existing semantic segmentation framework of U-Net neural network and combined with mineral liberation analyzer （MLA）, an automatic image particle analysis system was developed. By exploiting the backscattered electron （BSE） contrast hierarchy （Bi>steel matrix>MnS>silicate） and using MLA data as training masks, the system enabled precise classification of MnS, elemental Bi, and their binary/ternary composite inclusions. Furthermore, by statistically comparing errors with BSE results, an inclusion learning preference was construted to correct errors caused by MLA resolution limitations, thereby improving model accuracy. The system outputs multidimensional quantitative data including inclusion type, size, area, equivalent diameter, and spatial distribution. Statistical analysis reveals that increasing Bi content slightly improves the clustering of inclusions but also promotes their coarsening. Both number and size of elemental Bi particles increase significantly with higher Bi content. Moreover, Bi tends to exist in larger inclusions in the form of Bi-MnS composites. This methodology is not only applicable to current Bi-containing inclusion systems but can also be extended to the intelligent recognition and quantitative analysis of other composite second phases, providing a reliable technical pathway for quantitative characterization and performance optimization of secondary phases in steels.