Quality control is essential for product consistency and reliability in modern manufacturing. This work presents an end-to-end computer-vision pipeline that automatically detects and classifies perforation defects in industrial imagery. A two-stage architecture is proposed: a YOLOv11 detector localizes individual perforation regions, and a multi-label EfficientNetB0 classifier assigns one of four defect classes (DT0 non-defective, DT1 missing, DT2 touching, DT3 out-of-bounds). To address the scarcity of labeled data, we combine active learning and bootstrapping to select representative images for manual annotation, then apply an adaptive-thresholding pseudo-labeling stage to automatically label the selected samples, expanding the training set from a small hand-labeled subset to a substantially larger validated dataset. The final classifier reaches F1-micro 99.08% on held-out test data, demonstrating that the framework can support process monitoring and preventive maintenance.