Computer vision module manufacturing: optics, calibration and supply for embedded cameras

What makes a vision module hard to manufacture

A computer vision module packages one or more image sensors, lenses and often a processor on a small board that another company builds into a robot, kiosk, appliance, drone or inspection machine. It takes two common forms: a camera module that streams images over MIPI CSI-2, USB or a serializer link to a host processor, and a smart module that runs inference itself and sends results. Either way, manufacturing is hard for three reasons: optical alignment at the scale of microns, per-unit calibration that the host’s algorithms rely on, and keeping both stable through temperature, vibration and years of use.

A person looking at a consumer camera will forgive small differences in focus. An algorithm calibrated and trained on one module may lose accuracy on the next. So the specification has to define optical acceptance in measurable terms, such as sharpness by field position, distortion, optical center and tilt tolerance, and calibration error, and the factory has to measure every unit against it. If your specification says “good image quality”, that is the first thing to fix.

What the optical specification should contain

Each line below needs a limit, a test method and a test condition (lighting, distance, temperature) that the factory and your incoming inspection both use:

Parameter	How it is usually measured	What to state
Sharpness	SFR or MTF on a slanted-edge chart	Minimum value at a stated spatial frequency, at the center and at defined field positions
Working distance	Sharpness at near and far distances	The distance range the algorithm needs
Distortion	Grid or checkerboard chart	A maximum, or a requirement that calibration models it
Optical center and tilt	Chart alignment	Tolerance in pixels from the sensor center, and maximum tilt
Shading	Flat-field light source	Maximum corner fall-off, before or after correction
Defective pixels	Dark and flat-field frames	Maximum count and rules for clusters
Calibration quality	Reprojection error on a target; depth error for stereo	Maximum error at set distances

Module types and what each asks of the factory

Module type	Typical use	Manufacturing focus
Single RGB camera	Kiosks, appliances, inspection	Focus, cleanliness, image quality limits
Global-shutter camera	Robotics, barcode reading, fast motion	Sensor sourcing, trigger and lighting synchronization
Stereo depth	Mobile robots, obstacle detection	A rigid baseline; per-unit stereo calibration and its stability
Time-of-flight or structured light	Close-range 3D sensing	Illuminator eye safety; depth calibration with temperature compensation
Smart module with NPU	On-device inference	Thermal design, model packaging, SDK support
Multi-camera with serializer links	Robots and vehicles with long cable runs	Connectors, cable EMC, frame synchronization

Key components and sourcing

• Image sensor. The longest-lead and highest-risk part. Check the sensor’s expected production life, allocation and minimum purchase through distribution, and whether a driver and ISP tuning exist for your customers’ host processors. ISP tuning is specific to each combination of sensor, lens and host SoC, so a module sold into several host platforms carries tuning work for each one.
• Lens. Off-the-shelf M12 lenses are quick to source but vary more from lens to lens; custom lenses need tooling and a longer qualification but give control over field of view, distortion and the filter. Choose an IR-cut, visible-only or dual-band filter to match the lighting.
• Lens holder and adhesive. Materials with low thermal expansion, and adhesives with known shrinkage and creep, decide whether focus and calibration hold after curing and over temperature.
• Board and interconnect. Rigid or rigid-flex, a flexible tail or a connector, and memory on the module for calibration data and the serial number. MIPI CSI-2 runs only short distances over flex; longer runs need serializer links and automotive-style connectors.
• Illumination. IR LEDs or VCSELs and their drivers, for active depth sensing or night operation.

Core manufacturing processes

1. Sensor attach. Packaged sensors are reflowed onto the board; chip-on-board designs mount and wire-bond the bare die, which needs a clean room and wire-bonding capability. Particles on the sensor or filter are a leading cause of optical rejects.
2. Lens alignment. Passive designs thread the lens to focus and then glue it. Active alignment powers the sensor, images a target and positions the lens in several axes before UV curing. It gives tighter focus and tilt across the field and is common for higher resolutions and wide-angle lenses, but alignment machines limit throughput.
3. Curing. A UV tack followed by a thermal cure. Focus can move during cure, so measure after it.
4. Calibration. Intrinsic parameters (focal length, principal point, distortion) are measured for each unit with charts or collimators. Stereo modules add extrinsic calibration between cameras; time-of-flight modules add depth, phase and temperature calibration. Results are written to the module with a version and checksum.
5. Final test. Sharpness across the field, shading, color, defective pixels and noise; interface stability at full frame rate; illumination output; depth accuracy at set distances.
6. Traceability. Every unit’s calibration and test data is stored against its serial number. Integrators often ask for it, and it is the evidence behind a first article inspection.

Testing and validation from EVT to PVT

Stage	What to prove for a vision module
EVT	Optical performance against the specification across the field and temperature range; signal integrity over the real flex or cable length; integration with the customer’s host processor and algorithm; depth accuracy
DVT	Thermal cycling, vibration and shock, re-measuring focus and calibration after each; humidity and lens fogging; ESD and EMC at the connector; eye-safety classification for active illumination; packaging protection
PVT	Alignment yield; focus and calibration capability (Cpk) against limits; repeatability of the calibration station (gauge R&R); correlation between the factory test and the customer’s incoming inspection

Calibration drift is the failure that reaches customers. A stereo module that meets its depth specification at the factory and drifts after a week on a vibrating robot has a design problem, and DVT is the time to find it.

Compliance for a component

A module built into another product is not usually certified as an end product, but integrators will ask for:

• RoHS and REACH declarations and material data.
• Eye-safety classification for active illumination: IEC 60825-1 for lasers, including VCSELs, and IEC 62471 for LEDs. In the US, laser products are also subject to FDA rules for radiation-emitting products, so agree with the integrator who classifies and reports the finished product.
• Pre-compliance EMC data, so the integrator’s product is less likely to fail its own testing.
• Modular radio approval if the module carries a radio.
• Component-origin documentation where the end customer has procurement rules, such as NDAA requirements in US government and drone programs.
• For automotive programs, the component qualification and factory quality-system requirements of that industry.

Main cost drivers

• The image sensor, especially global-shutter, large-format or low-light parts.
• The lens: glass elements, custom designs and their tooling.
• Active alignment and calibration time per unit.
• Clean-room assembly and optical yield.
• Serializers, connectors and cable assemblies for long-reach designs.
• Driver and ISP tuning work for each host platform.
• Variant count: each field-of-view or filter option multiplies tooling, calibration and inventory, each at lower volume.

What to ask a prospective OEM/ODM

• Which camera modules have you built with active alignment, and what focus and tilt capability can you show?
• What clean-room class applies to sensor and lens assembly?
• How do you calibrate, with which targets and software, and what is the station’s repeatability?
• Where is calibration data stored, in what format, and who owns the station software?
• Which sensors have you sourced, through which channel, and how do you handle end-of-life notices?
• Can you support driver and ISP tuning on our customers’ host processors?
• Which reliability tests do you run, and do you re-measure calibration afterward?

Common mistakes

• Writing image quality requirements that cannot be measured.
• Calibrating at room temperature and assuming the result holds.
• Selecting a sensor late in its life cycle.
• Forgetting that ISP tuning has to be repeated for each host platform.
• Leaving the calibration format and ownership unstated.
• Ignoring flex-cable bend life and connector retention in moving machines.

Where Zefinity helps

We help teams define measurable optical and calibration requirements, find and qualify module makers with the alignment, calibration and clean-room capability the product needs, and follow DVT and pilot builds through NPI and production launch. For a complete camera product rather than a module, see the edge AI camera guide. See our work in AI hardware and industrial and robotics.

Frequently asked questions

What is active alignment in camera module manufacturing?

Active alignment positions the lens relative to the image sensor while the sensor is powered and imaging a target, adjusting in several axes until sharpness across the field meets the limit, then fixing the lens with UV-cured adhesive. It holds tighter focus and tilt than threading a lens and gluing it, at the cost of specialized equipment and longer cycle time.

How often does a stereo camera need recalibration?

Ideally never in the field, which is why the mechanical design matters so much. Calibration drifts when the baseline bends, adhesives creep or lenses shift with temperature, shock and vibration. Test calibration stability in DVT by measuring reprojection or depth error after thermal cycling and vibration, and design a field check or recalibration routine if drift remains.

Who should own a vision module's calibration data?

Agree it in the contract. Calibration parameters, their format, the station software and the raw test records are part of the product. Store the parameters in the module with a version number and checksum, and keep a copy keyed to the serial number in your own systems, so any module can be traced or replaced without depending on one factory.