<p>Advanced high-strength steels (AHSS) exhibit pronounced springback and shape variability, posing persistent challenges for quality control in free-form bending (FFB) processes. This study presents a real-time, inline inspection and automated classification framework that couples high-speed laser line scanning with principled feature construction and small-sample learning. Cross-sectional profiles are acquired at line rate and transformed by Principal Component Analysis (PCA), which preserves dominant geometric structure while reducing dimensionality. Within the PCA space, critical deviations are localized using a Ramer–Douglas–Peucker–based greedy segmentation strategy and projected onto a discriminative two-dimensional plane anchored to the principal trend of acceptable parts. On this plane, a margin-based Support Vector Machine with a radial basis kernel learns a robust decision boundary between acceptable and defective products. To mitigate optimistic bias arising from limited data, model selection and performance estimation are conducted via k-fold cross-validation on the labeled cohort, and generalization is further verified on held-out cases against expert judgment. The resulting pipeline enables reliable, low-latency, and non-contact quality decisions, providing a practical foundation for smart manufacturing and facilitating future integration into closed-loop, adaptive FFB control.</p>