Matlab Use Visual Inspection

Visual Inspection with MATLAB

Build machine vision inspection systems using Visual Inspection Toolbox. Select the right approach — rule-based, AI-based, or hybrid — based on problem characteristics, then implement, train, and deploy.

When to Use

Defect detection and classification

• Detecting anomalies on manufactured parts (scratches, dents, contamination)
• Training anomaly detectors with good-part images only (one-class learning)
• Detecting and localizing defects with bounding boxes (YOLOX)
• Classifying defect types or severity levels

Measurement and gauging

• Measuring widths, gaps, angles, diameters on parts
• Calibrated measurement in physical units (mm, degrees)
• Vision fixturing — locating parts and repositioning measurement tools
• Blob analysis — area, perimeter, shape metrics of segmented regions

Counting and localization

• Counting identical rigid parts via shape matching
• Few-shot counting with intra-class variation (CounTR)
• Locating parts by edge geometry for downstream inspection

Training data augmentation

• Generating synthetic labeled data via copy-paste augmentation
• Expanding small defect datasets for YOLOX training

Deployment

• Packaging inspection as standalone executables (MATLAB Compiler)
• Generating C/C++ for embedded/edge targets (MATLAB Coder)
• GPU-accelerated inference with TensorRT (GPU Coder)
• ONNX export for NPUs, AI accelerators, smart camera SDKs
• Deploying to Beckhoff TwinCAT industrial PLCs

When NOT to Use

• Camera acquisition, frame grabber configuration — use Image Acquisition Toolbox
• PLC communication, OPC UA, MQTT protocols — use Industrial Communication Toolbox
• General image processing unrelated to inspection (artistic filters, medical imaging)

Must-Follow Rules

Approach Selection

• Rule-based for geometric, deterministic problems — when defects are dimensional deviations or parts have known geometry, use caliper/fitcircle/matchshape. No training data needed
• AI-based for variable appearance — when defect appearance can't be enumerated, use anomaly detection or YOLOX. Requires training data
• Hybrid when parts move — use matchshape to fixture (locate + normalize pose), then AI on the cropped/warped region. Anomaly detectors are NOT rotation/scale invariant

Anomaly Detection

• Anomaly detectors require consistent pose — CNN backbones are not rotation/scale invariant. Fixture with matchshape + imwarp to canonical orientation before training and inference
• Choose detector by data availability — >100 good images: studentTeacherAnomalyDetector; <100 images: patchCoreAnomalyDetector; have anomaly labels: fcddAnomalyDetector
• Use classify for inference — returns [tf, score, map] for anomaly map visualization

Object Detection (YOLOX)

• Use AutoResize=false for full-resolution inference — prevents downscaling that loses small defects. Pair with tiled training when objects are small relative to image
• Synthetic data benefits YOLOX most — copy-paste augmentation via objectInsertionDatastore is proven for detection tasks

Measurement

• Vision fixturing pattern — matchshape locates part → transform measurement positions to detected pose → measure at repositioned locations
• Calibrated measurements require imageToWorldPlane — converts pixel measurements to physical units using camera calibration
• images.geotrans.Warper for production rectification — precompute the mapping once, apply per frame for maximum throughput

Deployment

• Always use persistent + isempty guard in codegen entry points — load model once, infer many times
• coder.loadDeepLearningNetwork for DL models — not coder.load (which is for structs/shape models)
• coder.Constant for MAT file path — codegen requires compile-time constant paths
• Pass detector objects directly to exportONNXNetwork — never extract .Network or .Backbone first. The object-level export includes pre/post-processing logic
• ONNX for NPUs/AI accelerators; GPU Coder + TensorRT for NVIDIA — don't use ONNX for Jetson/Orin

Pipeline Construction

Machine vision inspections compose tools in sequence:

1. Acquire — Get the image (Image Acquisition Toolbox)
2. Fixture — Locate a stable reference feature (matchshape)
3. Reposition — Transform measurement positions to detected pose
4. Inspect — Apply measurement or classification at repositioned locations
5. Decide — Compare to tolerances → pass/fail
6. Communicate — Send results to PLC/SCADA (Industrial Communication Toolbox)

Common Hybrid Patterns

Primary Decision: Rule-Based vs. AI-Based

Key Functions

Decision Trees

"I need to detect defects"

"I need to measure parts"

"I need to count objects"

"I need to locate parts"

"I don't have enough training data"

"I need to deploy"

Patterns

Each reference file contains executable code patterns, API calling conventions, and worked examples for its topic. Load the relevant reference before writing code.

Conventions

• Always: use machine vision terminology (rule-based, tool chaining, fixturing)
• Always: recommend the specific VIT function, not generic IPT building blocks
• Always: consider hybrid pipelines (rule-based fixturing + AI classification)
• Always: ensure anomaly detector receives images in consistent pose
• Always: use persistent + isempty for model loading in codegen entry points
• Always: pass detector objects directly to exportONNXNetwork (never extract .Network)
• Never: recommend AI when the problem is purely dimensional measurement
• Never: recommend rule-based when defect appearance is highly variable
• Never: use coder.load for deep learning models (use coder.loadDeepLearningNetwork)
• Never: use ONNX export for NVIDIA targets (use GPU Coder + TensorRT)
• Prefer: anomaly detection when only good-part images are available
• Prefer: YOLOX when labeled defect examples with bounding boxes exist
• Prefer: shapemodel/matchshape over YOLOX for locating rigid parts
• Prefer: images.geotrans.Warper for production rectification throughput
• Prefer: MATLAB Compiler for dashboards and UI-based monitoring apps

References

Copyright 2026 The MathWorks, Inc.