🤖 AI

Data Collection for Training Quality-Control AI in Carpet Manufacturing

This paper presents a comprehensive blueprint for an in-line machine-vision system in woven carpet manufacturing that integrates real-time defect inspection with a structured data collection and annotation strategy to systematically train and improve quality-control AI models, thereby addressing production bottlenecks and reducing defect rates within a Six Sigma framework.

Original authors: Akbar Erkinov

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Akbar Erkinov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a factory that makes giant, beautiful carpets. These carpets are like long, continuous ribbons of fabric, moving incredibly fast down a production line. For years, the only way to check if the carpet is perfect has been to have human workers stand by the line, squinting at the moving fabric, or to cut off small pieces to inspect later.

The problem? Humans get tired, they miss tiny flaws, and they can't keep up with the speed of modern machines. If the factory decides to make more carpets by adding new weaving machines, the human inspectors become a traffic jam. They simply can't check every inch, so bad carpets slip through the cracks, costing the company a lot of money.

This paper proposes a solution: A "Smart Eye" system that doesn't just look, but learns.

Here is the breakdown of their plan, using simple analogies:

1. The Problem: The "Bottleneck"

The factory is about to speed up production (like adding more lanes to a highway). But the "inspection lane" (the human workers) stays the same size. This creates a bottleneck.

The Risk: If they don't fix this, thousands of defective carpets will reach customers. The paper estimates this could cost the company 50,000 Euros every single day.
The Goal: They want to catch every flaw, not just the obvious ones, and do it fast enough to keep up with the machines.

2. The Hardware: The "Super-Camera"

Instead of a regular camera, they propose a specialized system:

Line-Scan Cameras: Imagine a camera that doesn't take a photo of a whole scene at once, but instead scans the carpet like a laser beam, line by line, as it zooms by.
Two Types of Light: Just like you might use a flashlight to see a stain on a shirt (bright light) or look at the side of a table to see a scratch (grazing light), this system uses two lights at once. One reveals color stains, and the other reveals physical bumps, holes, or broken threads.
The Stitch: Since the carpet is wider than a single camera can see, they use several cameras side-by-side and stitch the images together in software, like a panoramic photo.

3. The "Cold Start" Problem: How do you teach a computer what a defect is?

This is the cleverest part of the paper. Usually, to teach an AI to find defects, you need thousands of photos of defects. But in a factory, defects are rare. You might go weeks without seeing a specific type of flaw. You can't wait months to collect enough "bad" photos to train the AI.

The Solution: The "Annotation Flywheel"
Think of this as a self-improving loop:

Day 1 (The Safety Net): The system starts by learning what "perfect" looks like. It only sees good carpet. It learns the normal patterns so well that if it sees anything weird, it flags it as a "suspicious anomaly." It doesn't know what the flaw is yet, just that it's not normal.
The Human Helper: When the AI flags something weird, it shows it to a human inspector. The human says, "Ah, that's a broken thread," or "That's a stain."
The Flywheel: The system saves that specific image and the human's label. Now, the AI has learned a new type of defect.
Repeat: As the factory runs, the AI gets better and better. It starts recognizing the flaws on its own, only asking humans for help when it's truly confused. The more it runs, the smarter it gets, creating a "virtuous cycle" where the data collection is the product.

4. The Strategy: Growing Up

The paper suggests the AI matures in stages, like a student:

Stage 1 (The Novice): Just looks for "anything that isn't normal" (Unsupervised).
Stage 2 (The Student): Starts learning to name the flaws (e.g., "That's a stain," "That's a hole") once it has seen enough examples.
Stage 3 (The Expert): Can draw a precise outline around the flaw to tell the machine exactly how much of the carpet needs to be cut out or fixed.

5. The Result: The "Six Sigma" Goal

The paper ties this back to a business concept called Six Sigma, which is a way of measuring quality.

Before: The factory had a high rate of errors (about 20,000 defects per million opportunities).
After: By catching almost every flaw instantly, the number of "escaped" bad carpets drops dramatically.
The Payoff: This isn't just about quality; it's about money. By catching the flaws before the carpet gets glued and finished, the factory stops wasting money on carpets they know are going to be thrown away.

Summary

The paper proposes building a camera system that acts as a continuous, tireless inspector. Its main innovation isn't just the cameras, but the method of learning: instead of waiting for a perfect dataset to exist before starting, the system builds its own dataset while it works, turning every hour of production into a lesson that makes the next hour of production better. It turns the factory itself into a teacher for the AI.

Technical Summary: Data Collection for Training Quality-Control AI in Carpet Manufacturing

Problem Statement
The paper addresses a critical bottleneck in woven carpet manufacturing: the inability of traditional manual visual inspection to scale with modern production speeds and widths. As throughput increases, the gap between production volume and inspected surface area widens, leading to inconsistent quality control, subjective judgments, and missed defects. This issue was framed within a specific Six Sigma (DMAIC) project at a woven-carpet facility anticipating a 40% increase in output (from $7.0 \times 10^5$ to $1.0 \times 10^6$ m²). The facility faced a baseline Defects Per Million Opportunities (DPMO) of approximately $2.0 \times 10^4$ (a sigma level of 2.33) and a financial exposure of roughly €50,000 per day if the inspection bottleneck remained unaddressed. The core challenge is that high-quality industrial AI models require large, labeled datasets of rare and diverse defects, which do not exist on day one of deployment.

Methodology
The authors propose an end-to-end engineering blueprint for an in-line machine-vision system designed not only for real-time inspection but primarily to systematically collect and label defect data. The methodology is structured around four key technical pillars:

Imaging and Acquisition Subsystem:
- Hardware: The system utilizes synchronized line-scan cameras (specifically three stitched 8k sensors) to cover a 4-meter web width without motion blur.
- Illumination: A dual-illumination head captures the web under two conditions: bright-field (for tonal faults like stains and shade variations) and grazing/dark-field (for structural faults like broken picks, floats, and pile crush).
- Resolution Budget: The system is engineered to resolve features as small as 0.5 mm with a 3-pixel margin, requiring a line rate of 6 kHz at a web speed of 1 m/s. This generates a raw data rate of approximately $1.44 \times 10^8$ pixels per second per channel.
Defect Taxonomy:
- A practical taxonomy is defined, categorizing defects by their origin in the "weaving-to-gluing" process. Classes include Structural (e.g., broken picks), Pile (e.g., matting), Colour/Dye (e.g., shade bands), Contamination (e.g., oil marks), Holes/Tears, Pattern (e.g., misalignment), and Backing (e.g., delamination).
- Each class is mapped to the most effective illumination channel and the appropriate model type (Detection, Segmentation, or Anomaly Scoring).
The "Annotation Flywheel" (Data Collection Strategy):
- Recognizing the "cold-start" problem (lack of labeled defects), the authors propose a staged, self-reinforcing data collection process:
  - Stage 0 (Bootstrap): Immediate deployment of unsupervised anomaly detection trained solely on defect-free material (using the carpet category of the MVTec Anomaly Detection benchmark).
  - Stage 1 (Mining): The unsupervised model flags potential anomalies, which are saved as candidates.
  - Stage 2 (Human-in-the-Loop): Inspectors review only these high-probability candidates, confirming faults and assigning labels. This drastically reduces annotation costs compared to exhaustive frame-by-frame labeling.
  - Stage 3 (Augmentation): Rare classes are balanced using geometric/photometric augmentation and defect synthesis (compositing real defects onto normal backgrounds).
  - Stage 4 (Re-training): Periodic re-training of supervised models on the growing dataset, with active learning prioritizing uncertain cases.
Staged Modelling Strategy:
- The system evolves from an Unsupervised Anomaly Detector (available Day 1) to Supervised Classification (early stage), then Supervised Detection (mid-stage, for bounding boxes), and finally Supervised Segmentation (mature stage, for pixel-precise maps of area-extended defects). The anomaly detector acts as a permanent safety net for novel faults.

Results and Performance Projections
The paper is a design proposal and does not report measured experimental results from a deployed pilot. Instead, it provides engineering budgets and conceptual projections:

Throughput: The proposed hardware is calculated to handle the target 1 m/s speed with sufficient resolution to detect 0.5 mm defects.
Sigma Improvement: The authors argue directionally that replacing sampled manual inspection with 100% automated coverage will reduce the "escaped defect" count. If the system catches a fraction $r$ of previously missed defects, the DPMO should decrease proportionally to $(1-r)$ , theoretically moving the process toward higher sigma levels.
Economic Impact: By catching defects early (at the weaving stage rather than after gluing), the system prevents value addition to condemned material, directly addressing the €50,000/day financial exposure.

Key Contributions

Data-Centric Design: The primary contribution is treating data collection as a "first-class engineering goal." The system is explicitly designed to generate the labeled dataset required for future model improvement, rather than relying on pre-existing data.
Dual-Illumination Architecture: A specific hardware configuration combining bright-field and grazing light to address the diverse visibility regimes of carpet defects.
The Annotation Flywheel: A concrete workflow that transitions a manufacturing line from unsupervised operation to supervised excellence, minimizing the initial labor burden while continuously improving model capability.
Six Sigma Integration: The proposal maps technical components directly to DMAIC phases (Define, Measure, Analyse, Improve, Control), demonstrating how AI can serve as the instrument for process capability improvement.

Significance and Claims
The paper claims to offer a "deployable blueprint" for AI-assisted quality control in carpet manufacturing. Its significance lies in bridging the gap between theoretical deep learning capabilities and the practical realities of industrial deployment, specifically the lack of initial defect data. By grounding the proposal in a real-world Six Sigma project with defined financial and throughput constraints, the authors argue that camera-based inspection is the necessary lever to prevent production bottlenecks. The paper modestly acknowledges that it is a design proposal, not an evaluated deployment, and notes that future work requires pilot testing to measure actual capture rates and false-alarm rates against human baselines.