Changing the Sensor Means Recalculating the Illumination

LIGHTING

A quantum efficiency of 15 % at 940 nm asks for 4.9 times more light in photons and 2.85 times in watts, and for a moving part that multiplier triples again.

For near-infrared inspection of a black connector housing, I recalculated lighting, lens and WD from the published quantum efficiency of a BSI global shutter sensor.

About 16 min read

When the colors of print and paint cover the background in visible light and bury defects, moving to the near infrared (NIR)¹ is a common next move. Around 940 nm many organic dyes absorb far less, so the colors fade and the structure remains. The plausible idea that comes along is that switching to a new, more sensitive sensor lets the lighting stay as it is.

The Nexora family of BSI global shutter sensors that Teledyne e2v introduced on September 22, 2026 offers a chance to test that idea in numbers, because its published quantum efficiency (QE)² is up to 73 % at 550 nm and 15 % at 940 nm. Change only the wavelength and leave the lighting as it is, and the image goes dark; lengthen the exposure and moving parts blur; raise the gain and noise covers the defects. At the end of that road, loosening the threshold adds missed defects and tightening it adds false alarms.

What I want to work out from this number is what 15 % demands of the lighting design. Once it is known how many times more is needed in photons, in watts and in exposure time, the lighting power supply, the lens and the WD follow. The numbers use Nexora’s published values, and any value not published is stated as an assumption.

Black connectors split two ways at 940 nm

What to look at before the sensor is how the surface to be inspected reflects and transmits at 940 nm. What creates usable contrast in the near infrared is the material’s difference in reflectance by wavelength, and two areas that were different colors in visible light sometimes converge to the same reflectance in the near infrared, making the contrast vanish instead.

A black connector housing is a good example. If the black comes from a carbon black³ pigment, it still absorbs strongly at 940 nm and stays dark, while an organic black dye transmits more in the near infrared, so fine cracks or embedded particles beneath may show through. The same black connector can give opposite images depending on the pigment family. When lots with different gloss or surface treatments are mixed in, reflection changes too, so which way it goes cannot be guaranteed before a sample test.

So before changing the wavelength, I shoot a few real samples under 940 nm lighting, one image each in reflection and transmission. Only after confirming that the defects show do the numbers for sensor and lighting carry meaning. Two images, one in reflection and one in transmission, usually also reveal which pigment family is involved.

Fifteen percent is the absorption of about 9 μm of silicon

A BSI (backside illumination)⁴ structure flips the wafer to put the photodiode on the side where light enters, so the share of light lost passing between metal wiring shrinks structurally. Even so, the reason it stays at 15 % at 940 nm lies in silicon itself. The absorption depth⁵ at which silicon reduces light to 1/e is 1.56 μm at 550 nm but 54.6 μm at 940 nm (Green 2008 data).

If the silicon layer of a pixel is 3 μm, 85 % of 550 nm light is absorbed but only 5.3 % of 940 nm light. Conversely, absorbing 15 % requires light to travel about 8.9 μm through silicon, and 50 % requires 37.9 μm. Getting 15 % out of a thin layer takes a structure that lengthens the path light travels inside the silicon, so the fight to raise near-infrared sensitivity is in the end a fight over path length.

Pixel structures that lengthen that path also appear in patents. The OmniVision patents in the table below place near-infrared enhancement structures that bend and scatter light on the illuminated face of the photodiode, or put a reflective layer on the opposite side to hold the redirected light inside its own pixel. In place of thicker silicon, they make light travel back and forth within the same thickness.

Patent	Title	Assignee	Priority · granted
US 9,991,309 B1	CMOS image sensor having enhanced near infrared quantum efficiency	OmniVision Technologies, Inc.	2017-07-05 · 2018-06-05
US 10,224,364 B2	CMOS image sensor having enhanced near infrared quantum efficiency and modulation transfer function	OmniVision Technologies, Inc.	2017-07-05 · 2019-03-05
US 10,964,744 B1	Light control for improved near infrared sensitivity and channel separation	OmniVision Technologies, Inc.	2019-09-13 · 2021-03-30
US 10,587,830 B2	Method for controlling an active pixel image sensor	Teledyne e2v Semiconductors SAS	2015-08-03 · 2020-03-10

The specification of the first patent reports a simulation with two or more near-infrared enhancement structures per pixel in which quantum efficiency at 940 nm rose from about 11 % to about 34 %. Converted to absorption depth, 11 % corresponds to about 6.4 μm of silicon and 34 % to about 22.7 μm, so the path grew about 3.6 times. The second patent sizes the reflective layer so the redirected light does not leak into neighboring pixels and erode sharpness, and the third stacks filter layers to handle near-infrared sensitivity and color channel separation together. The last one, from Teledyne e2v, deals with the memory node of a global shutter and comes back where the exposure is shortened.

In photons 4.9 times, in watts 2.85 times

With the same sensor, lens and exposure, the number of signal electrons is the number of incoming photons multiplied by the quantum efficiency. When quantum efficiency drops from 73 % to 15 %, getting the same signal takes 73 ÷ 15 = 4.87 times, about 4.9 times, the photons. The figure is a concept diagram of how this ratio turns into a lighting multiplier.

The multipliers in the figure are counted in photons. Lighting power is bought in watts, so one more conversion is needed: the energy of a photon is inversely proportional to its wavelength, so the same 1 W delivers 1.71 times more photons at 940 nm than at 550 nm. Calculated in watts, the required irradiance⁶ is 4.87 ÷ 1.71 = 2.85 times. That is why I write near-infrared lighting specifications with the photon multiplier and the watt multiplier on separate lines.

The habit of measuring in lux with a light meter also has to stop here. Lux is a unit weighted by the sensitivity of the human eye, so 940 nm light reads close to 0 lx however strong it is. Near-infrared lighting has to be measured in W/m² or in photons, and matching the electrical input as well means checking the power conversion efficiency of the two LEDs separately in their datasheets.

Shorter exposure sends the lighting to strobes

When the part moves, the numbers are multiplied once more. At 10 μm/px, which puts a 30 μm crack on three pixels, a part passing at 50 mm per second (assumed) smears across three pixels in the 0.6 ms exposure used in visible light, and cutting the blur to one pixel takes 0.2 ms, a third. The photon multiplier then becomes 14.6 and the watt multiplier 8.5. The 14.7 in the figure is 4.9, rounded, multiplied by 3.

Carrying this multiplier with continuous lighting breaks heat and power capacity first, so the lighting moves to strobes fired short and hard. Firing 0.2 ms pulses 30 times a second keeps the light on 0.6 % of the time; average output and heat drop by this duty⁷, and only the peak output rises. The lighting power supply and heat sinking have to be sized for the peak, more than the average, to avoid falling short under pulsed drive.

Strobes presuppose a global shutter⁸. With a rolling shutter, where each row has a different exposure time, a short pulse lands on only some rows and creates banding. The last Teledyne e2v patent in the table above has the global shutter’s memory node hold the charge at the end of exposure for correlated double sampling⁹, while making the memory capacity at least N times that of the photodiode and transferring charge N times during exposure to widen the dynamic range under strong light. That direction is welcome to anyone firing strobes hard.

The three numbers show that a move to the near infrared is closer to replacing the lighting power supply than to replacing the sensor. It asks for 4.9 times in photons but 2.85 times in watts, and packed into strobes, the average load drops to the 0.6 % of time the light is on.

The C-mount lens sets the WD and the focus

Nexora comes in 12 MP (4,096 × 3,072) and 16 MP (4,096 × 4,096), and its 23 × 21.5 mm package was announced as fitting 29 × 29 mm cameras and standard C-mount optics. At 10 μm/px, which puts a 30 μm crack on three pixels, 12 MP gives a 41.0 × 30.7 mm field of view and 16 MP gives 41.0 × 41.0 mm. The pixel pitch is not in the announcement; taking it as 3.0 μm (assumed) gives 0.3 magnification and a WD¹⁰ of about 217 mm with a 50 mm lens, while a 35 mm lens gives about 152 mm, tight for fitting the strobe heads.

The C-mount is a 17.526 mm flange focal distance¹¹ standard with many off-the-shelf lenses, but the lens image circle¹² has to cover the sensor diagonal. Calculated at a 3.0 μm pixel, the 12 MP diagonal is 15.4 mm and the 16 MP diagonal 17.4 mm, so the 16 MP square format narrows the choice to lenses that support a large diagonal. In the near infrared, a lens whose chromatic aberration is corrected only for visible light shifts its focal plane, so focus set in visible light is off at 940 nm. My rule is to choose a lens rated for near-infrared correction and to refocus under 940 nm lighting.

Setup table for the move to 940 nm

Gathering the calculations into one table shows that the specification for a move to the near infrared is set in the lighting cells more than in the sensor cells. The first row also records that the crack, the particle and the unfilled area all share the same 10 μm/px.

Item	Value	Condition · basis	In plain words
① Minimum defect size (μm)	Housing crack 30 μm wide · 300 μm long, internal particle 80 μm, unfilled adhesive 0.2 mm²	At 10 μm/px: crack on 3 pixels, particle on 8, unfilled area on 2,000	The crack width sets the pixel size
② Lighting	940 nm near infrared, ±25 nm bandwidth, 0.2 ms strobe × 30 per second (0.6 % duty)	4.9 times in photons (14.6 at a third of the exposure), 2.85 times in watts (8.5). Cannot be measured in lux	Size the power supply for peak output
② Sensor · lens · WD	BSI global shutter 12 MP (4,096 × 3,072) · 16 MP (4,096 × 4,096), published QE 73 % (550 nm) · 15 % (940 nm)	With a 3.0 μm pitch assumed, a WD of about 217 mm with a 50 mm lens must be secured (thin-lens approximation), diagonals 15.4 · 17.4 mm, near-infrared corrected lens	Check the lens image circle and focus at 940 nm
③ Algorithm parameters	Secure signal with irradiance more than gain, separate flat-field taken at 940 nm	Threshold 3σ (re-estimated from 30 good parts after the wavelength change)	Do not reuse the visible-light threshold

The cell that governs the design in this table is the multiplier in ②. Sensor, lens and algorithm are choices, but 4.9 and 2.85 are lower bounds that come from physics, so the right order is to set the lighting power capacity and heat sinking and fit the rest inside them. Lens transmission and bandpass filter losses multiply on top of these lower bounds.

The opposite choice has its own territory. If visible light already gives enough contrast, a move to the near infrared only inflates the lighting budget, and if the part is stationary with time to spare in the tact, a long exposure with continuous lighting and a rolling shutter is more economical. If absorption beyond 1,100 nm has to be seen, that lies outside the limit of silicon and an InGaAs-based SWIR camera fits, and in extreme low light where a few photons per pixel have to be counted, a photon-counting sensor is the better route over improved quantum efficiency. Even at the same wavelength, results change with the pigment family, surface treatment and lot variation, so which way fits cannot be guaranteed before a sample test.

Choosing the sensor is the vision engineer’s job, but the lighting cells of this table have to be set together with whoever designs the lighting power supply and heat sinking. Whoever approves the sensor swap should check whether the quote also includes the lighting power supply and a near-infrared corrected lens. This article explains technology and is not investment advice. See our Disclaimer.

Field Note

To move to 940 nm, I would read the values at both wavelengths from the datasheet quantum efficiency curve, in place of simply swapping the wavelength under the same lighting, and write both the photon and the watt multiplier on the first line of the lighting specification. Next I would place a real connector under 940 nm lighting, refocus, and measure anew how many pixels the crack covers and the good-part σ after the wavelength change. Where the part is moving, I would look at pulse width and blur in the same table.

Field Checkpoints

Once the multipliers and the focus have been checked, finish with the items below.

• Is a WD of about 217 mm from lens to part (50 mm lens, 3.0 μm pitch assumed) actually available, clear of the strobe heads and fixtures? (recalculated with the confirmed pixel pitch)
• Have real samples been shot in reflection and transmission at 940 nm? (carbon black and organic dyes come out opposite)
• Has the lighting multiplier been calculated separately in photons and in watts? (4.9 and 2.85, or 14.6 and 8.5 at a third of the exposure)
• Was the near-infrared lighting measured in W/m² in place of lux?
• Is the lens rated for near-infrared correction, and does its image circle cover the sensor diagonal (17.4 mm for 16 MP)?
• Is the shutter a global shutter synchronized with the strobe? (a rolling shutter creates banding)
• Has σ been re-estimated from 30 good parts after the wavelength change?
• Are the lighting power supply and heat sinking designed for peak output?

Frequently Asked Questions

Would going to 850 nm lighten the lighting load?

It would. Silicon’s absorption depth is shorter at 850 nm than at 940 nm, so quantum efficiency tends to be higher for the same structure, and in the OmniVision patent’s simulation too, a pixel without enhancement structures reached about 15 % at 850 nm and about 11 % at 940 nm. Some organic dyes still absorb at 850 nm, though, and may not fade as much as at 940 nm. The wavelength is a value chosen with samples, weighing the lighting budget against the material’s transmission.

Can the gain be raised to match the brightness?

Raising the gain leaves the number of photons reaching the pixel unchanged, so the limit set by shot noise does not move. When photons are 4.9 times short at 940 nm, gain only brightens the picture, and the separation between defect and background returns only when the irradiance goes up. The flat-field and the threshold also have to be redone after the gain changes.

Glossary

• [1] Near infrared NIR
  Light beyond visible red, roughly between 700 nm and 1,100 nm. It is the infrared band that a silicon sensor can still receive. ↩
• [2] Quantum efficiency QE
  The share of photons entering a pixel that become signal electrons. At 15 %, 100 photons yield 15 electrons. ↩
• [3] Carbon black carbon black
  A black pigment made of carbon particles. It absorbs broadly from visible light into the near infrared and stays black at 940 nm. ↩
• [4] BSI backside illumination
  A sensor structure that flips the wafer to take light through the back side, which has no wiring layer. Less light is lost passing between wires. ↩
• [5] Absorption depth absorption depth
  The depth light travels in a material before falling to about 37 % (1/e) of its initial intensity. In silicon, 940 nm light has to travel 54.6 μm to fall that far. ↩
• [6] Irradiance irradiance
  The intensity of light reaching a surface measured in watts (W/m²). Unlike illuminance (lx), which is weighted by eye sensitivity, it works unchanged in the near infrared. ↩
• [7] Duty cycle duty cycle
  The share of each repetition period during which the light is on. Switching on for 0.2 ms 30 times a second gives 0.6 %. ↩
• [8] Global shutter global shutter
  A sensor mode in which every pixel starts and ends its exposure at the same instant. Even a short strobe lands evenly across the whole frame. ↩
• [9] Correlated double sampling CDS
  Reading a pixel once in its reset state and once after the signal has built up, and using the difference as the signal. It removes the pixel-to-pixel wobble of the reference level. ↩
• [10] WD working distance
  The distance from the lens to the surface being inspected. The value in the text uses a thin-lens approximation measured from the lens center. ↩
• [11] Flange focal distance flange focal distance
  The distance from the lens mount face to the sensor plane. For the C-mount it is fixed at 17.526 mm. ↩
• [12] Image circle image circle
  The diameter of the round area over which a lens forms a sharp image. If it is smaller than the sensor diagonal, the corners darken or blur. ↩

References

• Teledyne e2v, Teledyne e2v Introduces Nexora, Its Next Generation Family of Backside Illuminated Global Shutter CMOS Image Sensors (Business Wire press release, Seville: 12 MP · 16 MP, QE up to 73 % at 550 nm · 15 % at 940 nm, 3D stacked, LVDS, 23 × 21.5 mm package, C-mount)
• engineering.com, Teledyne e2v launches Nexora CMOS image sensor family (September 22, 2026)
• PVEducation, Optical Properties of Silicon (data of M. A. Green 2008; absorption coefficient 6.39 × 10³ /cm at 550 nm, 1.83 × 10² /cm at 940 nm)
• OmniVision Technologies, Inc., US patent US9991309B1, CMOS image sensor having enhanced near infrared quantum efficiency (priority and filing July 5, 2017; granted June 5, 2018)
• OmniVision Technologies, Inc., US patent US10224364B2, CMOS image sensor having enhanced near infrared quantum efficiency and modulation transfer function (priority July 5, 2017; filed August 1, 2017; granted March 5, 2019)
• OmniVision Technologies, Inc., US patent US10964744B1, Light control for improved near infrared sensitivity and channel separation (priority and filing September 13, 2019; granted March 30, 2021)
• Teledyne e2v Semiconductors SAS, US patent US10587830B2, Method for controlling an active pixel image sensor (priority August 3, 2015; filed July 29, 2016; granted March 10, 2020)
• Formulas: photon multiplier = QE(550) ÷ QE(940), watt multiplier = photon multiplier ÷ (940 ÷ 550), layer absorptance = 1 − exp(−α × thickness), required path = −ln(1 − absorptance) ÷ α, blur-limited exposure = pixel resolution ÷ travel speed, duty = pulse width × repetitions per second, field of view = pixel count × pixel resolution, WD ≈ f × (1 + 1/m), sensor diagonal = √(width² + height²) × pixel pitch

Related reading — An 850 nm Band-Pass Filter Erases Factory Lights but Leaves Sunlight · A Board at 0.5 m/s Leans 15 mm Under a Rolling Shutter · What Optical Design Needs to Not Miss Defects on a High-Speed Line