Machine Vision Quality Control: From Defect to Review
Machine vision supports product quality checks for features distinguishable in an image. A deployment needs a specific defect, suitable imaging and a clear action after an alert. In food production, examples include label position, completeness and visible packaging damage.
What an image can reveal
Start by listing observable deviations: a missing or misaligned label, an incorrectly fitted cap, a missing item in a set or a visible packaging defect. Define acceptable limits for each item and collect examples of good products, defects and borderline cases.
Visual inspection has a defined scope. Microbiological safety, composition and seal integrity require appropriate testing methods. A camera may identify a visible sign of damage; product suitability is decided under the established quality procedure.
Camera placement and lighting
Choose an inspection point where the relevant product surface is visible and there is time to act. Check the smallest defect's image size, line speed, camera distance, motion blur and occlusion by adjacent products.
Choose lighting for the surface and feature being inspected. Film glare, shadows and changing daylight can hide a defect or create a similar pattern. KEYENCE's guide explains how lighting arrangements reveal surfaces, outlines and irregularities. Test the choice on your packaging samples.
A general surveillance camera is suitable only for features that can actually be distinguished in its footage. Small defects and fast lines may require a dedicated camera, optics, lighting and synchronized capture. Establish this before estimating deployment costs.
Preparing validation data
Collect footage across shifts, batches and packaging variants. Note lighting changes, line changeovers and acceptable product variations. The quality team must interpret borderline examples consistently; otherwise evaluation will depend on who annotated each frame.
Reserve a separate batch or time period for final validation. Keep these frames out of model and threshold tuning to assess performance on new data. Agree on the evaluation unit in advance: an individual product, a defect or a batch.
Acceptance metrics
Record actual defects found by the system, missed defects and good products that triggered alerts. Recall measures the share of confirmed defects detected. Precision measures the share of system alerts that correspond to confirmed defects. Google's machine learning guide defines these metrics.
Measure alert latency and duplicate notifications for the same product separately. Break results down by defect type, batch and imaging conditions. When defects are rare, testing on a defect-free stream does not establish detection reliability.
Agree acceptance thresholds based on the consequences of misses and false alarms. Before testing, record sample size, manual-review method and the treatment of borderline cases.
What happens after detection
Define the assignee and available response time. An alert may trigger a clip review, an additional product or batch check, or a quality-log entry. Automatic rejection and equipment interaction require separate engineering and safety validation.
The log should help staff locate the product or episode, investigate the cause and record an outcome. Useful fields include time, zone, deviation type and an image. Agree the data with the quality team's existing workflow.
Where to start
Prepare a defect list, sample images, line speed, product dimensions and a description of current checks. Busted offers a free initial review and a paid 14-day pilot with an agreed scope. The pilot tests feasibility and produces measurements to support a deployment decision.