ISO 9001 Certified Manufacturer 360° Factory Tour
Applications

Selecting Filters for Machine Vision Inspection

· 6 views · 3 min read · OOTEE Optics Editorial

Original content by the OOTEE Optics editorial team, drawn from first-hand production line and inspection data. Please credit the source when republishing.

Given limits in expertise and test conditions, the theories, parameters and conclusions here carry no guarantee of absolute accuracy. For technical learning and exchange only - never the sole basis for design, acceptance or selection. Corrections welcome.

Selecting Filters for Machine Vision Inspection

What a filter contributes to inspection

A machine vision filter selects the wavelength, polarization state or intensity reaching the camera. Its purpose is stable separation between a defect and its background across parts and operating conditions. A monochrome sensor still has a wavelength-dependent response.

Start with representative good and defective parts. Establish useful lighting geometry and wavelength before choosing a filter. With an existing illuminator, measure its spectrum at operating temperature first. Filtering cannot create a scratch signal that the lighting geometry never produced.

Principles and parameters: evaluate spectral overlap

Center wavelength (CWL) locates a passband. Full width at half maximum (FWHM) is the distance between its half-peak edges, not a one-sided half-width. Optical density (OD) describes attenuation; angle of incidence (AOI) affects the position of an interference passband. A light-emitting diode (LED) needs a measured spectrum, not just a nominal wavelength.

For fractional transmittance $T$:

$$\mathrm{OD}(\lambda)=-\log_{10}T(\lambda)$$

OD3 means 0.1% transmission within the specified blocking range. A near-infrared leak can still reach a sensitive sensor even when visible blocking is adequate.

At fixed exposure, a linear, unsaturated pixel signal can be approximated by:

$$S_i=K\int E(\lambda)\rho_i(\lambda)T(\lambda)Q(\lambda)\,\mathrm{d}\lambda$$

Here $E$ is illumination at the part, $\rho_i$ its effective reflectance for the chosen geometry, and $Q$ camera response. $K$ includes exposure and imaging scale. Ambient illumination is a separate contribution. Similar defect and background spectra mean that spectral filtering may reject interference without increasing their intrinsic contrast.

Gaussian bandpasses at 630 nm with 90% peak transmission
Fig. 1 Gaussian bandpasses at 630 nm with 90% peak transmission

In Fig. 1, the LED has 20 nm FWHM. Both filters have CWL 630 nm and 90% peak transmission. Integration at 1 nm intervals gives approximately 40% LED energy throughput for the 10 nm filter and 75% for the 30 nm filter. These are mathematical examples; measured flat-top passbands require a fresh calculation.

Selection: separate spectral interference, glare and saturation

Compare defect separation under candidate illumination wavelengths, then establish the exposure budget. When specifying a bandpass filter, include LED bin variation, heating, CWL tolerance and the angles of marginal rays. Matching filter FWHM to LED FWHM does not transmit all useful light.

SymptomFirst approachVerify
Ambient light changes monochrome imagesBandpass matched to illuminationSignal margin across the field
Glare on glass or metalAdjust geometry, then test crossed polarizationDefect signal as well as glare
Overall saturationReduce light or exposure, then consider NDBright regions and faint defects

A trial specification might use CWL 630 nm, FWHM 30 nm, peak transmission at least 90%, and OD3 blocking over 400–600 nm and 660–1000 nm. Define AOI, clear aperture and transition regions before procurement; this is a starting point for testing.

Ask for transmission data across the sensor's useful range, including any secondary passbands. Compare the dimmest acceptable defect with the brightest normal surface at the required cycle time. A filter that produces a cleaner image at long exposure may leave too little signal for a moving production line.

Crossed polarization uses a source polarizer and a lens analyzer, initially 90° apart. Glass glare often responds well. Bare metal, curved parts and rough surfaces require testing because reflection can change polarization. If scratches fade with the glare, change geometry or relax the crossing angle.

A neutral density filter (ND) helps when exposure or aperture must stay fixed. OD0.6 transmits about 25%, reducing linear signal to roughly one quarter. It also attenuates faint defects rather than selectively removing glare. Check spectral flatness over the working band.

Application: a metal-scratch decision sequence

Bandpass filtering can stabilize monochrome mark inspection against ambient light. Glass inspection often starts with glare control. Metal scratches first need useful scattering contrast. Follow Fig. 2 while changing one main factor per trial.

Filter selection sequence for metal scratches
Fig. 2 Filter selection sequence for metal scratches
  1. Collect faint scratches, acceptable machining texture and different part poses. Fix cycle time and acceptance criteria.
  2. Test low-angle darkfield with 470 nm and 630 nm illumination. Select the combination that separates scratches from normal texture.
  3. Measure the warm LED spectrum. For a 630 nm LED with about 20 nm FWHM, start with a 30 nm bandpass and compare ambient lights on and off.
  4. Test crossed polarization if glare remains. Return to lighting adjustments if scratch information disappears.
  5. Reduce illumination or exposure before adding ND. Validate across field position, temperature limits and multiple part batches.

Common mistakes: judge detection performance

A narrower bandpass can waste photons and increase sensitivity to wavelength shifts. Deeper blocking cannot remove glare at the illumination wavelength. A rotating analyzer alone is also different from a complete crossed-polarization arrangement.

Hold exposure, gain and processing fixed during comparisons. Record saturation, defect separation and noise with automatic exposure disabled. Accept the system against missed defects, false rejects and cycle time on an agreed sample set, retaining raw images from the worst temperature, pose and ambient-light conditions.

Include scratches near the acceptance limit rather than testing only obvious damage. Rotate representative parts to expose directional machining marks and specular highlights. After any filter change, check focus and field uniformity again: additional glass and restricted apertures can affect imaging as well as brightness. Record the final mechanical arrangement so that subsequent assemblies reproduce the tested geometry. Repeat acceptance checks after replacing the illuminator with a different production batch.

Related Reading

View all

Related Products

View all products

Need a custom optical filter?

Send us your drawing and spectral targets — an engineer replies with a feasibility check and quotation within one business day.

Source Factory In-house coating lines, no middlemen
Ready to Ship Standard items dispatched same or next day
Quality Guaranteed 100 % inspected, spectral report included
Express Delivery DHL / FedEx worldwide