Building a Real-Time Defect Detector with Line-Scan + ML (Reusable Playbook)

TL;DR

Use line-scan for moving/continuous stuff: film, foil, paper, textiles, wire, pipes, board edges, etc.
Decide early: field of view (FOV), smallest defect, speed, working distance (WD).
Lens: pick focal length so FOV at your WD matches product width; keep ≥3–5 px on the smallest defect.
Encoder on the motion path → square pixels and mm-accurate measurements.
Pipeline: camera → acquire → tile → preprocess → CV/ML → PLC/UI/logs. Start around f/5.6 (or f/8 if you need more depth of field).

When line-scan makes sense

Continuous webs (paper, plastic film, foil, textiles, glass, sheet metal).
Long parts (cables, rods, extrusions, tubes).
Narrow critical zones (seams, edges, labels).
Why line-scan: uniform lighting across width, effectively unlimited image length, high resolution along motion.

Optics: the 60-second recipe

FOV ↔ focal length

\text{FOV} \approx \frac{\text{Sensor Width} \times \text{WD}}{\text{Focal Length}}
\quad\Rightarrow\quad
\text{Focal Length} \approx \frac{\text{Sensor Width} \times \text{WD}}{\text{FOV}}

Ensure the lens image circle ≥ sensor (e.g., 1.1″ lens for a 1.1″ sensor).
Start aperture at f/5.6; go f/8 if the surface isn’t perfectly flat and your light is strong; open to f/4 if exposure is tight.

Resolution target: smallest defect should span ≥3–5 px in both axes.

Motion sync (why the encoder matters)

A rotary incremental encoder on a driven roller (or a measuring wheel on the web) outputs pulses as the product moves. Configure the camera so 1 pulse (or every N pulses) = 1 line.

Keeps mm/px constant (no stretch/squash if speed drifts).
Gives true distances in mm for defect size/position.
Stabilizes ML (consistent aspect ratio).

Back-of-napkin sizing: if your optics give mm/line = 0.15, then lines/mm ≈ 6.67. Pick encoder pulses/mm ≈ 6.67 (use a programmable PPR or a divider in camera).

Lighting (pick by defect type)

Transmissive backlight: holes/tears/missing material (silhouette).
Diffuse top light: stains/print/color issues.
Grazing/dark-field: scratches/raised fibers/embossing.
Add polarizers to cut glare. Always perform dark/flat-field correction.

Throughput sanity check

Lines/s (square pixels): LPS = belt_speed_mm_s / mm_per_line
Data rate (8-bit): MB/s = pixels_per_line × 1 × LPS / 1e6
1–5 GigE is plenty for most single-camera web lines; jump to 10 GigE/CXP if you push very high speeds or bit depths.

System architecture (Mermaid)

flowchart TD
  A["Moving product/web"] --> B["Backlight / Top light"]
  B --> C["Line-scan camera"]
  D["Rotary encoder on roller"] -->|"A/B pulses"| C
  C --> E["NIC / Frame grabber"]
  E --> F["Acquisition service"]
  F --> G["Tile builder\n(256–512 lines + overlap)"]
  G --> H["Preprocess\n(flat-field, denoise, contrast)"]
  H --> I["Detection"]
  I --> J["Classical CV"]
  I --> K["ML model (ONNX/TensorRT)"]
  J --> L["Decision"]
  K --> L
  L --> M["PLC / Marker / Alarms"]
  L --> N["UI dashboard"]
  L --> O["Logs / Database"]

Software pipeline (that holds up in production)

Acquire encoder-triggered lines (camera SDK; fixed IP; jumbo frames if Ethernet).
Tile: stack 256–512 lines with small overlap.
Preprocess: dark/flat-field, optional denoise/CLAHE.
Detect:
- Classical CV for structure/continuity/holes.
- ML (YOLO-style or small U-Net) for complex defects; export to ONNX → TensorRT.
Act: I/O to PLC/marker, save crops, stream overlays to a lightweight UI, log metrics.

Commissioning checklist

Set WD, square/level the camera, tension/stabilize the product path.
Focus at object plane; lock focus & iris.
Wire encoder (prefer RS-422 differential); enable termination; set camera TriggerSource=Encoder.
Do dark & flat-field calibration and save the gain map.
Verify mm/px with a ruler pass; tweak encoder divider/scale.
Seed known defects; tune thresholds and alert logic.
NIC: fixed IPs, MTU 9000, power-saving off.

Good defaults

Aperture: start f/5.6 (→ f/8 if you need more DOF).
Motion blur limit: keep motion during exposure ≤ 0.3 × mm/line.
Pixels on smallest defect: ≥3–5 px.
Encoder: incremental quadrature, programmable PPR if you can.

Typical single-lane BOM (sketch)

Line-scan camera (4K–8K) + C-mount lens matched to FOV/WD (image circle ≥ sensor).
Lighting (backlight/top) + driver + diffuser + optional polarizers.
Incremental encoder + mount/wheel + shielded cable.
Industrial PC (8-core CPU, 16–32 GB RAM, NVMe) + GPU if ML.
Matching NIC (1G/5G/10G) or CXP grabber; shielded cabling.
24 V I/O to PLC/markers; UPS; frame/mounts/enclosure.