Many instruments show their results only on a display. When those readings are needed from many instruments at once and at a high rate, manual note-taking becomes the weakest link. Industrial cameras can read displays automatically and send the numbers straight into a spreadsheet. Below is how this works and how we implemented display reading on a water meter test bench.
Why not manual notes
- Speed. A person cannot write down the readings of several instruments every second; a camera can.
- Errors. Copying numbers introduces mistakes: swapped digits, a skipped line, a reading from the wrong instrument.
- Linking. Every reading is automatically assigned to a specific instrument and a specific moment in time.
- Operator time. The operator supervises the process instead of sitting with a notepad.
- Traceability. Results go straight into a spreadsheet, so they are easy to review, compare and archive.
Why OCR on LCD displays is not trivial
Industrial OCR is different from scanning documents. LCD displays have their own difficulties:
- seven-segment characters, which standard OCR tools recognise poorly;
- reflections from the display glass and contrast that depends on the viewing angle;
- readings change, so a frame may catch the display mid-update;
- the display shows not only numbers but also status icons, which also have to be recognised and evaluated.
That is why everything matters together: the choice of camera and lens, camera position, lighting and a recognition model trained specifically for these displays. For more on choosing cameras, see industrial cameras.
Case study: water meter testing on a metrology bench
On a metrology test bench, 4 Basler dart USB3 cameras are mounted above the meters. Once per second they read the LCD of every water meter. The system is in production.
- What is read. Volume, flow rate and status icons: valve, radio, leak and battery.
- Identification. Each meter is identified by its QR code, so a reading is always assigned to the correct instrument.
- Comparison. The readings are compared with the reference values.
- Results. Data is written to a spreadsheet (Excel).
- Software. Python, OpenCV and pypylon (the Basler camera library). Characters are recognised by a CNN model we trained ourselves. An earlier version of the system ran on Aurora Vision (Adaptive Vision) Studio.
- Operator panel. A web HMI (Node.js). The panel for 12 meters was accepted in October 2026.
- Mechanics. The cameras use 3D-printed housings.
The key point is that the system does not read just one meter but every meter on the bench. Each reading is linked to a specific meter by its QR code, so the operator does not have to match records to instruments by hand. Status icons are checked together with the numbers: the test covers not only volume and flow rate but also what the instrument reports about its own state.
What CNN character recognition is
A CNN (convolutional neural network) is a type of neural network for images that learns from examples. Instead of describing by hand what each seven-segment digit looks like, the model is trained on many real photos of the display with the correct answers. Such a model is tailored to a specific display type, font and set of icons. If the instrument model changes, it can be retrained with new examples without changing the whole system.
How such a project runs
- Trial with real instruments. We photograph your displays under real conditions: actual lighting, distance and angle.
- Choosing cameras and lighting. We work out how many instruments one camera can see and what resolution is needed.
- Recognition model. We prepare and train character and icon recognition for your displays.
- Integration. Results go to a spreadsheet, database or PLC, and the operator sees them on the panel.
- Commissioning and verification. We compare the readings with manual or reference values until the operator is confident the system reads correctly.
Where else display reading fits
Meter testing is just one example. Display reading makes sense wherever an instrument's result is visible only on its display, and connecting to it through a data interface is impossible, too expensive or would take too long. This is often older equipment, test benches and inspection operations where many identical instruments are tested at once. For the wider context, see quality inspection with machine vision, and for neural networks applied to images, AI segmentation in manufacturing.
We start display reading and meter testing projects with a trial on your real instruments. Find out more in our machine vision and AI services and projects. Do you have a test bench where readings are still written down by hand? Get in touch.