An industrial camera costing a few hundred dollars, plus recognition software running on a small computer in the workshop, can do work that used to require two people watching a line all shift. The question is not whether the technology works. It is which problems are worth spending money on.
Which problems fit
Not everything should be automated with a camera. A good candidate has four characteristics, and if one is missing you should think twice:
- It repeats many times a day. Inspecting 3,000 items a shift is worth it. Inspecting 20 is cheaper done by a person.
- People make mistakes when tired. Watching continuously for eight hours, the miss rate rises noticeably towards the end of a shift. Machines do not tire.
- It can be judged by eye. If an experienced person can tell by looking, a model can learn it. If it requires taste, smell or touch, a camera will not solve it.
- Errors cause measurable loss. Wrong shipments, returns, contract penalties — you need a number to calculate payback.
Packing and warehousing
Verifying contents before sealing
A camera over the conveyor compares what is in the box against the order before the lid closes. Missing items, extra items and wrong variants get caught on the spot rather than by the customer.
The value lies in the fact that handling one incorrect shipment usually costs several times the value of the item: return shipping both ways, customer service time, and marketplace reputation.
Reading barcodes and labels in bulk
One camera reads many codes across a whole pallet instead of scanning box by box. For warehouses with high throughput, this is where time saving is most visible.
Checking packaging quality
Crushed boxes, misaligned labels, lids not fully closed, torn shrink wrap. These are the defects a human inspector tends to miss at the end of a shift.
Agriculture and food processing
Grading produce by size, colour and defect
This is the most mature application. Grading tomatoes, mangoes, dragon fruit or coffee beans by size and ripeness. A machine grades to the same standard all shift, whereas human graders tend to relax the standard as the day goes on.
For export produce, consistent grading against the buyer's specification directly affects the price achieved and the rejection rate.
Counting and monitoring livestock
Counting animals in a pen, spotting one lying abnormally, flagging an animal separated from the group. Done manually this is both laborious and error-prone.
Detecting disease on leaves
Cameras mounted on equipment moving through a greenhouse, spotting disease markers earlier than the human eye. Harder than produce grading because it needs far more training data, but valuable at scale.
Other industries
| Industry | Problem | What the camera does |
|---|---|---|
| Metal fabrication | Surface inspection | Detects scratches, pitting, missing drill holes |
| Textiles | Fabric inspection | Finds yarn defects and colour drift between rolls |
| Construction | Site safety | Flags missing helmets, entry into restricted zones |
| Retail | Shelf management | Detects empty shelves and misplaced stock |
| Logistics | Yard management | Reads plates, logs entry and exit automatically |
| Healthcare | Instrument counting | Reconciles surgical instruments before and after a procedure |
What it actually costs
The items below are what a single line involves, so you know what to expect before requesting a quote. Actual cost depends on conveyor speed, how hard the visual task is, and lighting conditions in your facility.
| Item | Note |
|---|---|
| Industrial camera and lens | Selected for conveyor speed and object size |
| Dedicated lighting | Often underestimated, but stable light affects accuracy more than the camera does |
| On-site processing computer | Runs in the facility, independent of internet connectivity |
| Data collection and labelling | Needs several thousand sample images from your own line |
| Model development and tuning | Where most of the time goes |
| Installation and calibration | Real conditions never match the lab |
Payback is fastest on lines running multiple shifts with high volume and expensive escapes. Slowest on single-shift lines with low volume and frequently changing products.
How to validate before committing
Do not sign for a whole-factory rollout. Pilot one line and one product type for four to six weeks. The pilot costs a fraction of full deployment and answers the only question that matters: at your actual lighting and line speed, what accuracy do you achieve.
Three things that decide success
Lighting matters more than the camera
This surprises people new to the field. A mid-range camera with proper dedicated lighting outperforms an expensive camera under light that changes through the day. If budget is tight, spend it on lighting first.
Training data must come from your own line
A model trained on stock images will perform poorly on your actual product, under your facility's light, at your camera angle. There is no shortcut here.
Decide up front what happens when the machine is wrong
No system is 100% accurate. The question to settle at the start is: when the model is uncertain, should it reject the item or route it to a person. For export food, rejecting a good item beats letting a bad one through. For high-value goods, routing to a human review makes more sense. This decision belongs to the business, not to engineering.
Where to start
Pick one step where somebody currently inspects by eye. Count how many items pass through per day and estimate the current escape rate. Those two numbers are enough to know whether the problem is worth solving, before any technology is discussed.
Written from the team's hands-on experience delivering infrastructure and software. No sponsorship, no product placement.