Computer Vision QC for Industrial Manufacturer

FabricaPro's quality control was the bottleneck. Eight human inspectors reviewed fabric throughout the day, catching 88% of defects on average. That missing 12% hit customers as complaints, returns, and reputation damage. The economics didn't work either—rehashing faulty fabric cost more than just replacing it.

Designing the Vision System

We installed high-resolution cameras (20 megapixel) at inspection checkpoints and captured imagery as fabric moved through the line. For training data, we worked with FabricaPro's QC team to annotate 500 images with all types of defects: holes, missing stitches, color irregularities, and sizing issues.

We used a YOLOv8-based object detection model fine-tuned on fabric defects. The model learned to localize defects at pixel-level, which meant we could show inspectors exactly where to look—crucial for human-in-the-loop validation.

Deployment & Human-AI Workflow

The system doesn't replace inspectors; it augments them. When the CV model detects possible defects, it routes that fabric batch to a human inspector with annotations showing exactly where problems are. Humans make the final call on whether it's truly defective or a false positive.

This hybrid approach achieved 99.2% accuracy on defect detection while keeping false negatives below 1%. More importantly, it reduced inspector cognitive load—they're confirming predictions, not doing blind search.

Results & Scaling

Within 3 months of deployment, customer complaints dropped 87%. Manufacturing waste dropped 35%. The system processed 150+ tons of fabric daily with zero false negatives making it to customers (validated through post-shipment tracking).

FabricaPro is now rolling out the system across all production lines. The investment (€50K system + €20K implementation) paid for itself in <3 months through waste reduction and complaint costs avoided.