AI Quality Inspection

Deploy computer vision systems that detect defects faster and more consistently than human inspectors — running 24/7 on high-speed production lines.

Why AI for Quality Inspection?

Manual visual inspection is subjective, fatiguing, and cannot scale. AI vision systems offer:

• Consistency: Same detection criteria applied to every single product, 24/7
• Speed: Inspect hundreds of parts per minute at full production line speed
• Sensitivity: Detect sub-millimeter defects invisible to the human eye
• Data: Every inspection is logged with images and measurements for traceability

Inspection Types

Defect Detection Pipeline

import cv2
import torch
from ultralytics import YOLO

class DefectInspector:
    def __init__(self, model_path='defect_model.pt'):
        self.model = YOLO(model_path)
        self.defect_classes = ['scratch', 'dent', 'crack',
                               'stain', 'missing_part']

    def inspect(self, image_path):
        """Run inspection on a product image."""
        results = self.model(image_path, conf=0.3)
        defects = []

        for det in results[0].boxes:
            defect = {
                'type': self.defect_classes[int(det.cls)],
                'confidence': float(det.conf),
                'location': det.xyxy[0].tolist(),
                'area': self._compute_area(det.xyxy[0])
            }
            defects.append(defect)

        verdict = 'PASS' if len(defects) == 0 else 'FAIL'
        return {'verdict': verdict, 'defects': defects}

    def _compute_area(self, bbox):
        return (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])

# Run inline inspection
inspector = DefectInspector()
result = inspector.inspect('product_001.jpg')
print(f"Verdict: {result['verdict']}, Defects: {len(result['defects'])}")

Anomaly Detection for Unseen Defects

When you only have "good" examples, anomaly detection finds deviations from normal without labeled defect data:

• Autoencoders: Train on good images. High reconstruction error indicates anomaly.
• PatchCore: State-of-the-art industrial anomaly detection using memory bank of normal features.
• Student-Teacher: A student network trained on good data fails to reconstruct anomalous regions.
• MVTec AD: Benchmark dataset for industrial anomaly detection with 15 categories.

Hardware Setup