Manufacturing teams have a lot to watch on the factory floor. Even a small defect can slip through inspection, while a misplaced part can hold up the next step.
Computer vision in manufacturing uses images to support these everyday decisions. The important part is how the result fits your process. A flagged defect, for example, needs a clear route to review or rejection.
This guide shows you six manufacturing use cases. For each one, you’ll see how the image result supports a production decision and what to test on your own line.
What is computer vision in manufacturing?
Computer vision in manufacturing uses images or video to support a factory task. For example, it might flag a defect for review or locate a part so equipment knows where to act.
To do this, the system captures an image and checks the features relevant to the task. It then sends the result to a person or another system. That result will go to a programmable logic controller (PLC), which controls industrial equipment. In a sorting setup, the PLC uses the signal to divert a part from the conveyor belt.
In industrial settings, you’ll also see the two terms machine vision and computer vision overlap with each other. While computer vision analyzes images or video to extract useful information, machine vision uses image analysis to inspect products or guide equipment using programmed checks or trained models.
For a broader introduction, see our explanation of computer vision. On your production line, start with two questions: what does the image reveal, and what insights will the results reveal to you?
6 computer vision use cases in manufacturing at a glance
Computer vision applications in manufacturing serve different production decisions:
| Use case | Result and next step | When it fits | Main limitation | What to measure |
| Quality inspection | Flags a possible defect for review or rejection | When you have a visible defect with clear acceptance rules | Defects that are hidden or hard to see in the image | Missed defects and false rejects |
| Assembly verification | Flags an assembly issue for a second check | When you have clearly visible parts and clear rules for correct assembly | Obscured components or uncontrolled variants | Missed errors and unnecessary stops |
| Robot guidance and sorting | Uses the part’s position or orientation to guide handling | When you have clearly visible objects and a defined task for moving or sorting them | Calibration or timing mismatch | Correct picks or diversions |
| Code reading and traceability | Reads a code and updates the matching part record | When you know your code format and a clear destination for the data | Unreadable codes or incorrect associations | Codes read correctly and matched to the right parts |
| Production-flow monitoring | Event data prompts investigation | When you define station events and a response owner | Ambiguous boundaries or incomplete tracking | Event accuracy and timing error |
| Safety monitoring | Detected event prompts an alert | When you want to test a response to an observable event | Obscured views or excessive false alerts | Missed events and response time |
1. Automate visual quality inspection
Visual inspection fits when the camera shows the defect clearly, and you have a common rule for accepting the product. For example, detecting a surface scratch requires a clear view of the damaged area. You should also include a reference image that indicates what the surface looks like.
That makes lighting an important part of the setup. Light at a shallow angle, for instance, makes some scratches stand out. To measure a part’s size, you also need suitable optics and calibration to relate image measurements to real dimensions.
The person who decides what to do with the result also plays an important part. An operator who reviews a flagged part has a chance to check it before removal. With automatic rejection, the equipment acts directly on the vision result. In either case, there is still a chance that a missed defect lets a faulty part pass, or a false reject removes a good one. Set acceptable limits for both errors.
Test good and faulty samples under the conditions your station will face. Keep the scope clear, since a surface check cannot rule out every hidden internal fault.
2. Verify parts and assembly steps
Assembly verification checks for assembly mistakes before work moves on. For example, a camera might check a connector’s orientation before the operator closes its housing. This gives you a chance to catch the error while the part is still visible.
The evidence you need depends on what you are checking. To confirm that assembly steps happened in the right order, you need images of those steps over time.
Once you’ve defined the check, decide how to respond when it fails. At a manual station, an alert to the operator might be enough. On a controlled line, define the response the controller should trigger. Also plan for blocked camera views that leave the system unable to judge a part.
Next, test known assembly mistakes alongside correctly assembled product variants. Track missed errors separately from unnecessary stops. Use the results to see whether the system catches mistakes without interrupting correct work. During these tests, focus on assembly issues, such as missing parts or steps performed in the wrong order. Use quality inspection to assess surface finish.
3. Guide robots and sort materials
Consider a robot picking a part from a tray. Before it moves, it needs to identify the right part and know where to reach. Recognition identifies the part, while localization finds its position. If the pick also depends on which way the part faces, the system needs pose information.
To act on that information, the robot needs coordinates it understands. Calibration connects positions in the image to locations in the physical work area. Keep the camera and lens setup consistent with that calibration, and recalibrate when a change affects the mapping.
A moving conveyor adds a timing challenge because the part changes position between image capture and the equipment’s response. The output must therefore reach the equipment in time to guide the action. Control requirements matter too, since a robot pick and a PLC-controlled reject mechanism use different controls.
If you want to apply this to your business, follow the whole process from image to action. Record whether each pick or diversion was correct, then compare the outcome with the image result. This helps you distinguish an incorrect location from a delayed control signal and fix the stage that caused the failure.
4. Read codes and track parts through production
Code reading helps you track a part through production, but the value must stay linked to the right part and record. A serial number might be read correctly yet saved against the wrong unit. If that happens, the next process could use incorrect information.
Choose a reading method that matches the marking on the part. Barcode and QR readers decode symbols. Optical character recognition (OCR) reads printed characters, such as a batch number. Image quality matters too, so test with images from your station.
Once the code is read, decide where its value should go. For example, a reader might send a serial number to a manufacturing execution system (MES), which tracks production work and records. Define how that number will be matched to the correct unit. Then decide how to handle unreadable codes or the same event arriving twice.
Finally, test your record and the code reading. Check that the record belongs to the right unit and that no duplicate entries were created. If a read fails, confirm that the record follows your handling process before the next step uses it.
5. Monitor production flow and cycle times
Production monitoring uses visible events to measure how work moves through a line. Start by defining what the camera should count. For example, specify where a tray enters a station and where it leaves so every observation uses the same boundaries.
At a workstation, three common measures show different aspects of performance:
- Production throughput. Measures output per unit of time, so you can compare production rates.
- Dwell time. Records how long an item stays in a defined area, including any waiting there.
- Cycle time. Measures the time to complete a defined process cycle. Agree on its start and end events before comparing results.
Those measures show how work moves through the station but not the delay. If a tray stays longer than expected, check for waiting or rework. Assign someone to review the event against the station’s operating records. Before you rely on those measures, compare counts and durations with a separate reference sample. Missed events and unclear boundaries distort the results.
If a machine starts moving unusually, investigate its condition. That movement alone does not mean it will fail. To use the change for predictive maintenance, compare past observations with operating and maintenance records to see whether it preceded later faults. Then test whether the pattern helps predict failures in new cases.
Fun fact: Video-based activity tracking also has many use cases outside manufacturing. Our AI-driven in-store operations intelligence project used existing CCTV to track staff activity and supply dashboard metrics in the retail environment.
6. Monitor safety events on the factory floor
Computer vision for manufacturing safety needs a specific event to detect and a clear plan for responding. For example, checking whether someone enters a restricted zone gives you an observable event to test. A broad request to identify unsafe behavior is harder to define and measure.
Before collecting test samples, decide what must be visible. To check personal protective equipment (PPE), the camera needs a clear view of the relevant item. For a zone alert, define the boundary and make sure the camera can see someone crossing it. If the view is blocked, the system may be unable to confirm whether the event occurred.
Next, assign someone to review alerts and agree on when to escalate them. Give that person a clear process for handling uncertain results. During testing, measure both missed events and false alarms. Frequent false alarms add review work and can make useful warnings easier to overlook. Also decide who can access stored footage and how long it will be kept. Check these arrangements against your site’s privacy requirements.
Finally, keep the system’s safety role clear. Camera alerts support monitoring, but using them as safety-rated protective controls requires a separate engineering review. An alert system alone does not guarantee compliance or prevent every injury.
Computer vision in practice: Our conveyor sorting system
When equipment acts on a vision result, the detection must reach the controls in time. Our conveyor sorting project put that connection to work on a moving production line.
A US manufacturer needed to identify materials and locate their exact position on a conveyor belt. That position data had to trigger an existing PLC-controlled reject mechanism at the right moment, so it acted on the correct item.
To locate the materials, we used an object-detection model from the YOLO family. TensorRT sped up inference, which means applying the trained model to new images. This setup processed two camera feeds.
We ran the system on an edge graphics processing unit (GPU) connected to the existing controls. Those controls used the position data to time the rejection.
As a result, our computer vision sorting system case study reports 98% to 99% detection accuracy and a 40% to 60% latency reduction. For your pilot, record the image result alongside the action that followed. If the wrong item is diverted, compare those records to check two stages:
- Image result. Did the model identify the right material and locate it correctly?
- Control response. Did the signal reach the controls in time, and did the reject mechanism act on the correct item?
How to identify a potential computer vision project for your business
Start with a recurring production problem, then assess whether computer vision is practical and worth the cost with these three steps:
Define the task and cost of errors
First, describe the problem and document how you handle it today. For inspection, measure the time spent reviewing parts and the cost of missed defects or false rejects. Use those records to estimate the value of a change. Include any manual work that would remain after the system is introduced.
Next, define the result the system should produce. Plan how software systems integration will pass that result to the next step, and assign someone to act on it. With the workflow defined, include setup and running costs in the business case. This gives you an estimate for the whole task, including work beyond image analysis.
Check image quality and system integration
Once the task is clear, collect images of the product variants and conditions your pilot will face. Check whether the detail you need is consistently visible. If it is not, try a different camera view or lighting setup before treating the issue as a model-training problem.
Also define how the next step responds when a result cannot be used normally:
- Late or missing. Decide what the controller or operator should do when a result is missing or arrives too late.
- Uncertain output. Define a review or fallback process when the system lacks enough evidence for a decision.
- Repeated event. Check that receiving the same event twice does not trigger an extra action or create a duplicate record.
These response plans also depend on where image analysis runs. Choose local or cloud processing to match your timing and data needs. Then test what happens if the connection fails and confirm that the next step follows the agreed fallback.
Test production results before you expand
Before testing, agree on a safe setup with the people responsible for the equipment. Then compare the image results with your reference labels and verify what happens next. For traceability, check that your record is linked to the right part. For sorting, confirm that the equipment diverts the correct item.
During testing, log missed or incorrect detections separately from failures in the equipment’s response and data linked to the wrong part, so you can identify where each error occurred and which stage needs a fix. Also make clear who handles problems and who maintains the system.
Before expanding the rollout, agree on the evidence needed to justify that decision. Test across the shifts and product variants the system will support. One promising sample does not show that it will work reliably across all of them.
If your vision pilot needs a custom build, our AI integration services cover model testing and links to your existing systems. We agree on what to measure before the build. We set up monitoring for use after launch.
In conclusion
The value of computer vision in manufacturing depends on how you use its results on the production line. A useful pilot shows whether the next action works reliably and how much manual work or upkeep remains.
Start with one recurring problem and define success in a short pilot brief. Then test the complete workflow with the people who run it. Use the findings to weigh the measured benefit against setup costs and the effort needed to keep the system running. That gives you a clearer basis for deciding whether to expand.
