In brief
What makes a computer vision inspection project succeed?
Controlled, consistent imaging and training data that covers real variation. Fix lighting, camera position and resolution first; collect images across shifts, product variants and defect types; validate on data from different days; and design how uncertain results are reviewed by inspectors.
Visual inspection is one of the most practical uses of AI in manufacturing and logistics. It also has a reputation for pilots that work in the lab and struggle on the line. In our experience the reasons are rarely about algorithms. They are about the physical conditions in which images are captured and the data used to train and test.
Start with the image, not the model
A useful test is simple: can an experienced inspector reliably see the defect in the captured image? If the answer is no, no model will solve the problem reliably. Improve the imaging first.
| Factor | Common problem | Typical remedy |
|---|---|---|
| Lighting | Ambient light changes with time of day or shift; reflections hide defects | Enclosures, controlled diffuse or angled lighting suited to the defect |
| Camera position | Parts vary in orientation; defects appear at edges of the frame | Fixtures, guides or multiple camera angles |
| Resolution | Small defects span too few pixels | Higher resolution cameras or lenses matched to defect size |
| Motion | Blur on moving conveyors | Faster shutter, strobe lighting, triggering from sensors |
| Contamination | Dust or oil on lenses degrades images over time | Protective housings and cleaning routines |
Collect data across real variation
Training images taken over a few hours on one line capture a narrow slice of reality. Production brings different shifts, material batches, product variants, seasonal temperature and gradual equipment wear. Plan data collection to cover that variation deliberately.
- Capture images across multiple shifts, days and, where relevant, lines or sites.
- Include every product variant that will be inspected.
- Work with quality teams to define defect categories and labelling guidelines before labelling starts.
- Record metadata such as time, line and batch so performance can later be analysed by condition.
Dealing with rare defects
In a well-run process, defects are rare, which makes them hard to learn. Several approaches help, often in combination. Targeted collection gathers defect samples from quality records and rework areas. Anomaly detection learns what good parts look like and flags deviations. Data augmentation creates variations of existing defect images, used carefully so that artefacts do not mislead the model. Uncertain predictions go to inspectors, and their decisions become new training data.
Validate the way the line runs
Hold out data from different days or lines for testing rather than splitting images randomly. Measure both missed defects and false rejects, because each has a cost: escapes reach customers, while false rejects waste good product and erode inspector trust. Agree acceptable rates for each with quality leadership.
Design for the shop floor
An inspection system is used by people under production pressure. Alerts must be clear, uncertain cases must be quick to review and inspectors must be able to correct results easily. Plans are needed for camera cleaning, recalibration after maintenance and what happens if the system goes offline. Involving line and quality staff from the start makes these practical details far easier to get right.
About the author: Dr. AarzooCo-Founder and Chief Operating Officer, AARHIT Systems. Writes on delivery, quality assurance, data readiness, monitoring and human review. More articles by Dr. Aarzoo