AYXWORKS uses necessary cookies to work. Analytics and advertising cookies, including third-party cookies from Google, load only if you allow them. You can change this any time in “Cookie settings” at the bottom of the page.
Machine vision uses images or related visual measurements to inspect, identify, measure or guide a process. It does not always need AI. Stable geometry and well-controlled lighting may be handled effectively by conventional algorithms; learned models can help when appearance varies too much for practical hand-written rules.
Explanation, examples & sources
Define the decision first: defect present, dimension within tolerance, correct label or usable pick pose. Then test a simple baseline before adding model training. Keep lighting, camera placement and part presentation in scope; AI cannot reliably recover information that the image never captured. Measure false accepts and false rejects against the business and safety consequences. A classifier's overall accuracy is not enough when rare defects are the reason for installing the system.
Should an inspection or picking system use 2D or 3D vision?
Use 2D when the required information is visible in the image plane under controlled geometry, such as many labels, edges and surface checks. Use depth or 3D information when height, shape, volume or object pose is essential. The need comes from the measurement, not from 3D being a higher-numbered upgrade.
Explanation, examples & sources
A top-down label check may not need depth; picking overlapping parts from a bin often does. Test surface material, occlusion, field of view and required resolution at working distance. Allocate time for lighting, calibration and robot integration, not only sensor purchase. A 3D system can still struggle with transparent or reflective parts, and a simple 2D setup can outperform it on the right task. Validate the final error at the process output, not only the camera's nominal precision.
Why are lighting, calibration and coordinate frames critical in robot vision?
Lighting determines what the sensor can observe, calibration relates measurements to a model, and coordinate frames connect camera information to robot motion. Errors in any of these can make a visually convincing detection produce an incorrect measurement or movement. A high-confidence AI result does not correct a wrong geometric transform.
Explanation, examples & sources
Record camera intrinsics, distortion, camera-to-robot relationship, tool frame and the units used by every interface. Verify transform direction and timestamps with known targets before moving near equipment. Test the complete chain at multiple positions and orientations across the working volume. Recheck after changing the lens, camera mount, tool or fixture. Keep calibration data versioned with the hardware arrangement so a mechanical adjustment does not silently invalidate the saved configuration.
How should I validate a vision system before production?
Test it on representative and difficult cases that were not used to tune the system. Measure the errors that matter to the process, including false accepts, false rejects, uncertain results and failed acquisitions. Validate across expected variation in parts, lighting, positions and operating conditions rather than one curated image set.
Explanation, examples & sources
Split data so near-duplicate images of the same part or production batch do not leak between training and testing. Include damaged, borderline and unfamiliar examples with independently checked labels. Define threshold selection, escalation and drift monitoring before launch. For a guided robot, also measure pose and picking outcomes, not just image labels. Keep a versioned acceptance dataset and rerun it after model, optics or lighting changes. Where the vision output supports a protective function, ordinary model evaluation is not sufficient safety evidence.
How do detection, segmentation, classification and anomaly detection differ?
Classification assigns a label to an input; object detection locates instances, often with bounding boxes; segmentation labels image regions or pixels; anomaly detection flags departures from a learned or defined normal pattern. Choose the output that the process actually needs rather than using the most elaborate model available.
Explanation, examples & sources
A sorting gate may need only a category, while a gripper may need an object boundary and pose. Anomaly detection can be useful when defects are diverse, but a high anomaly score does not explain the defect or prove rejection is justified. Define annotation rules and ambiguous cases before collecting a large dataset. Evaluate the complete decision path, including thresholds and handling of uncertain outputs. More detailed predictions can increase annotation and validation work without improving the final decision.