Get Started with Visual Inspection

R2026b

Visual inspection is the process of examining manufactured parts, assemblies, or environments to identify defects, verify dimensions, and ensure compliance with quality standards. In industrial settings, visual inspection encompasses a range of tasks from detecting surface anomalies on production lines to measuring critical dimensions of machined components.

The Visual Inspection Toolbox™ provides algorithms and tools that enable you to automate visual inspection tasks such as:

• Quality control and defect detection — Identify anomalous regions in images of manufactured parts using deep learning models trained primarily on defect-free reference images.

• Dimensional metrology and tolerance verification — Measure distances, angles, and geometric features with sub-pixel accuracy using caliper and edge-fitting tools.

• Automated counting and inventory — Count objects in images using exemplar-based deep learning models that require only a few sample patches.

• Production line monitoring — Detect and localize objects using deep learning object detection models, with support for deployment to edge devices and hardware.

Visual Inspection Capabilities

This table summarizes the major visual inspection areas in the toolbox, with information on getting started and links to relevant functionality and examples.

Inspection Capability	Representative Image	Get Started	See Also
Anomaly Detection — Detect, classify, and localize defects in images using unsupervised deep learning models trained only on defect-free reference images.	For an example that uses this image, see Localize Industrial Defects Using PatchCore Anomaly Detector.	To interactively detect, classify, and analyze anomalies such as industrial defects, use the Visual Anomaly Detector app. You can also use the app to train, fine-tune, and evaluate anomaly detection deep learning models. You can also programmatically detect anomalies: To split training data, use splitAnomalyData. Train a Student-Teacher, FastFlow, PatchCore, or fully convolution data description (FCDD) anomaly detection network by using the trainStudentTeacherAnomalyDetector, trainFastFlowAnomalyDetector, trainPatchCoreAnomalyDetector, or trainFCDDAnomalyDetector function, respectively. To generate anomaly scores for new images, use predict. To classify images as normal or anomalous, use classify. Visualize results with anomalyMapOverlay and evaluate performance with evaluateAnomalyDetection.	Get Started with Anomaly Detection Using Deep Learning Localize Screen Manufacturing Defects Using Student-Teacher Anomaly Detector Localize Industrial Defects Using PatchCore Anomaly Detector Identify Defects in Air Compressors Using Spectrogram Images Detect Image Anomalies Using Explainable FCDD Network
Measurement — Measure distances, angles, and circular features in images using interactive and programmatic precision tools with edge-snapping accuracy.	For an example that uses this image, see Create an App to Measure Images Using Edge‑Snapping Metrology.	Measure distances between edges using caliper or interactively using uicaliper. Fit and measure circles to round features using fitcircle or interactively using uifitcircle. Measure angles between edges using fitangle or interactively using uifitangle. Refine positions to sub-pixel edge locations using snap. For calibrated cameras, map pixel coordinates to world units using imagePointsToWorldPlane.	Perform Metrology Edge Measurements and Alignment Create an App to Measure Images Using Edge‑Snapping Metrology Automate Calibrated Image Measurement Using Shape Matching
Shape Matching — Align inspection images to a reference using shape-based template matching. You can also detect and localize objects from a single template image without training data.	For an example that uses this image, see Detect Industrial Parts Using Multi-Class Shape-Based Matching.	Create a shape model from a template image using shapemodel. Search for the template in an inspection image using matchshape to recover object poses, consisting of translation, rotation, and scale. Preview edge features with show to tune extraction parameters. Use the recovered transformation to align measurements or annotate detections.	Detect Industrial Parts Using Multi-Class Shape-Based Matching Automate Calibrated Image Measurement Using Shape Matching
Object Detection — Detect and localize objects in images using YOLOX deep learning models, with support for generating synthetic training data.	For an example that uses this image, see Detect Defects on Printed Circuit Boards Using YOLOX Network.	Create or augment training data using insertObjectInImage and objectInsertionDatastore. Create and train a YOLOX detector using yoloxObjectDetector and trainYOLOXObjectDetector, respectively. Run detection on new images using detect. Export the trained model for deployment using exportONNXNetwork. You can also interactively detect objects in images and analyze detection results using YOLOX by using the Object Detector Analyzer app.	Detect Objects Using YOLOX Deep Learning Network Detect Small Objects Using Tiled Training of YOLOX Network Detect Defects on Printed Circuit Boards Using YOLOX Network Code Generation for Detect Defects on Printed Circuit Boards Using YOLOX Network
Object Counting — Count objects in images using only a few sample patches without requiring per-object annotation, by using exemplar-based deep learning models.	For an example that uses this image, see Count Objects Using CounTR Model.	Select exemplar patches of objects to count interactively using uiselectboxes or programmatically using extractpatches. Create a CounTR model for object counting using counTRObjectCounter. Count objects in new images using countObjects. Visualize the spatial distribution of counts using densityMap.	Count Objects Using CounTR Model

Representative Visual Inspection Procedure

A typical visual inspection procedure includes image acquisition and setup, inspection, and decision-making.

Image Acquisition and Setup

Acquire images from cameras or sensors and load them into the workspace. For an example of live image acquisition, see Detect Anomalies in Pills During Live Image Acquisition (Image Acquisition Toolbox). Use image datastores to manage large collections of inspection images and organize them into training, validation, and test sets for deep learning models. For more information on using datastores, see Datastores for Deep Learning (Deep Learning Toolbox).

Inspection

Perform the inspection task using one or more of the techniques described in Visual Inspection Capabilities. To select between various detection approaches based on your inspection goal and available data, see Select Detection Approach for Visual Inspection.

Decision Making and Results Communication

Evaluate inspection results against acceptance criteria to make pass or fail decisions. You can compare anomaly scores against optimized classification thresholds, verify that measured dimensions fall within tolerances, confirm that shape matching scores meet minimum confidence levels, or check that detected object counts match expected values. Assess overall system performance using confusion matrices, and metrics such as precision, recall, and F1 scores.

Communicate inspection results to automation systems for traceability or to trigger actions such as part rejection or line stoppage. Use OPC Unified Architecture (OPC UA) or other industrial communication protocols to send inspection outcomes to programmable logic controllers (PLCs). For more information, see Industrial Communication Toolbox™.

Deployment

Deploy trained models and inspection pipelines to production environments. Export models to the Open Neural Network Exchange (ONNX) format for integration with external inference frameworks. Generate optimized C or C++ code for CPU-based deployment, or CUDA^® code for NVIDIA^® GPUs. You can use MATLAB^® Compiler™ to package complete inspection applications as standalone executables for end users who do not have MATLAB. For more information, see Deploy Visual Inspection Code, Models, and Applications.