Moving Parts Are Frozen by How Long the Light Stays On

LIGHTING

What freezes a moving part is how long the light is on, more than the exposure time. At 1,000 mm/s, keeping blur under 0.5 px means the light must go off within 12.25 μs.

In a setup with a 60 mm field and 24.5 μm/px, I back-calculated the pulse width from the allowed blur, then worked out the limits of overdrive and how aperture and layout win back the lost light energy.

About 12 min read

When a jig carrying connectors passes under the camera at 1,000 mm/s, the metal pin tips turn into motion blur¹ dragged long in the travel direction. Shorten the exposure and the image darkens; raise the gain² because it is dark and noise starts to look like defects. Glossy pins and matte solder mask share one image, so matching both with a single exposure is hard as well. The compromise picked that way collapses again as soon as the line speeds up a little.

Blur stretches the outline of a defect along the travel direction and thins its contrast. Once the blur width nears the defect size, a pin-tip dent of about 80 μm mixes into the background and passes, and lowering the threshold to stop that floods the results with over-detected smeared normal edges. Either way, the inspection result becomes a value that sways with line speed.

The fix is to let the time the light is on stand in for the shutter. With a strobe³ that fires the light as a pulse a few μs wide, the image forms only for the pulse width even if the camera exposure window is longer. The light lost to the shorter time is made up by overdrive⁴, which drives the LED briefly at a current above its rating. I fixed a line with a 60 mm field at 1,000 mm/s and started the calculation from one line: blur px = speed × pulse width ÷ pixel resolution.

Blur is speed times the time the light is on

The length of motion blur is the travel speed of the part multiplied by the time light comes in. Divide that length by the pixel resolution and you get how many pixels it smears across in the image. At 0.5 px or less the steepness of an edge is hard to tell from standing still, and past 1 px the edge width grows noticeably. My rule is 0.5 px.

A 2,448 × 2,048 px sensor with a 16 mm F/5.6 lens viewing 60 mm across gives 24.5 μm per pixel (assumed). With continuous light and a 100 μs exposure, blur is 1,000 mm/s × 100 μs = 100 μm, 4.1 px, while the smallest defect of 80 μm is 3.3 px, so the blur is longer than the defect. Keeping blur under 0.5 px, or 12.25 μm, means the light must go off within 12.25 μs, and setting 10 μs for margin gives a blur of 10 μm, 0.41 px. Blur shrinks tenfold.

Open a film camera shutter at 1/60 s in a dark room and fire a flash, and the film records only the brief moment the flash passes. That is also why a spinning fan blade comes out looking frozen. A strobe works the same way: the camera exposure window only needs to wrap the pulse, and the pulse width does the real work of the shutter. So on the first line of a spec I would write the allowed blur in px ahead of the exposure time.

The current limit sits in one pulse, not average heat

Cut the pulse to 10 μs and light runs short, and overdrive makes up the shortfall with current. The heat limit of an LED is often described in terms of average power, peak power × duty cycle⁵. Taking 20 images a second, one every 50 mm of travel (10 mm overlap with a 60 mm field), the duty cycle of a 10 μs pulse is 10 μs × 20 = 200 μs/s, 0.02%. Even driven at four times the rated current, the average current is about 0.08% of rated continuous drive, so average heat is hardly an issue.

The constraints move elsewhere. They are the peak current that the LED chip and driver can take, the junction temperature rise during one pulse, and the current rise time⁶ of the driver. The SVS-Vistek technical notes also say that with short pulses an LED can be driven well above its rated current, but its lifespan may suffer, so the on-time should be as short as possible. On top of that, LED light output grows more slowly at high current, so four times the current does not guarantee four times the light.

With a driver whose rise time is 1 μs, 10% of a 10 μs pulse is spent while the current climbs, and the shorter the pulse, the larger that share. On the first day of setup I would capture the light current waveform on an oscilloscope to confirm how long light is actually emitted, and check the peak current, maximum pulse width and rise time, together with the duty cycle, against the pulse rating data of the manufacturer. Trusting the duty cycle alone comes back later as a slow failure, chip degradation.

Winning back the 60% of light energy that went missing

The light energy collected in one image is roughly proportional to light output × on-time. Switch from continuous light at 1× for 100 μs to a 4× overdrive pulse of 10 μs and the energy falls to (4 × 10) ÷ (1 × 100) = 0.4×, possibly lower in practice once LED efficiency loss is counted. Filling the gap with gain raises signal and noise together, so the contrast-to-noise ratio does not improve. Undoing the blur of a long exposure with software deblurring has the same problem of amplifying noise, so it cannot cover the shortfall of the lighting setup.

In my order of steps, the aperture comes first. Opening one stop from F/8 to F/5.6 gathers (8 ÷ 5.6)² = 2.04 times the light, bringing the energy back to 0.82×. The price is depth of field⁷. Taking the permissible circle of confusion as two pixels (6.9 μm), at this magnification the depth shrinks from 6.4 mm at F/8 to 4.5 mm at F/5.6, so whether the variation in pin height and jig height fits within 4.5 mm has to be checked as well.

The remaining 1.22× comes from moving the lights closer to the part or adding more of them, and gain stays as the last resort. Moving the lights closer means the bar lights and a shading hood have to fit inside the WD together, so a space calculation comes along with it. Writing the sharpness gained by narrowing the pulse and the light lost as energy ratios on the same table decides at the design stage whether the shortfall closes with the aperture or more lights have to be bought.

Seating a 10 μs pulse inside a 20 μs exposure window

The camera uses a global shutter⁸, in which every pixel starts and ends exposure at the same instant. The exposure window is set to 20 μs, wrapping the 10 μs pulse with margin on both sides. If the trigger jitter⁹, the wobble between the trigger signal and actual emission, is 1 μs, the part moves 1 μm, 0.04 px, in that time, which is negligible as a position error. With a short window, ambient light has only 20 μs to build up, yet blocking outside light with a shading hood keeps brightness even from frame to frame.

The figure below overlays long continuous light at rated current and a short 4× overdrive pulse on the same exposure window. The same edge is dragged more than 4 px under continuous light and breaks off within one pixel under the pulse.

Whether the calculation holds is checked at real speed. Pass a calibration pattern through at 1,000 mm/s, measure the width over which edge brightness rises from 10% to 90%, and see that it does not grow by more than 0.5 px over the stationary value. If it does, the tail of the current waveform is a likelier suspect than the pulse width, because when a driver lets current drain slowly at switch-off, light lingers long after the pulse.

From an 80 μm pin tip to a 20 μs window, one spec sheet

Moved into a spec, the calculation gathers the defects to catch, lighting and sync, the lens and WD (working distance)¹⁰, and blur verification with judgment criteria into one table. The WD is the distance from the front of the lens to the part, and the bar lights and the shading hood have to fit in between.

Item	Reference value	Condition and basis	In plain words
① Minimum defect size · pin-tip dent	80 μm or more	3.3 px at 24.5 μm/px. Design assumption	A dent about three pixels wide
① Minimum defect size · silk smear and solder ball	Silk smear 100 μm, solder ball 120 μm across or more	4.1 px, 4.9 px. Design assumption	Slightly larger smears and beads
② Optical setup · lighting	Two LED bar lights, 10 μs pulse, 4× rated overdrive	Confirm allowed current and pulse width with the pulse rating of the manufacturer	Switch on once, short and strong
② Optical setup · sync	20 images a second, 0.02% duty cycle, 20 μs exposure window, trigger jitter 1 μs or less	Jitter 1 μs × 1,000 mm/s = 1 μm (0.04 px)	Keep shutter and flash in time
② Optical setup · lens and sensor	16 mm F/5.6, global shutter 2,448 × 2,048 px, 24.5 μm/px	Magnification 0.141, thin-lens object distance 129.7 mm, depth 4.5 mm (2 px circle of confusion assumed)	60 mm across in one image
② Optical setup · WD	WD of 110 mm or more secured (design assumption)	From the lens front. Measure including interference from bar lights and hood	Lights and hood have to fit under the lens
③ Algorithm parameters · blur allowance	0.5 px or less (0.41 px at 10 μs)	4.1 px with 100 μs continuous light	Keep the drag under half a pixel
③ Algorithm parameters · blur check	Edge 10 to 90% width on the move grows 0.5 px or less over stationary	Measured by passing a calibration pattern at real speed	Compare running against standing still
③ Algorithm parameters · brightness and judgment	Per-frame gain correction from reference plate mean brightness (with a cap), over 3σ from background with area 6 px or more	An 80 μm defect covers about 10.7 px. Tracks light variation between pulses	Match brightness per flash, then look for flaws

Put the first line of the table beside the blur-allowance line and it shows where this configuration succeeds or fails. The smallest defect of 80 μm is 3.3 px, barely past the detection floor; the 4.1 px blur of 100 μs continuous light is larger than the defect and erases it in principle, while the 0.41 px blur of a 10 μs pulse is an eighth of the defect size. The price, light energy at 0.4×, comes back to 0.82× with one stop of aperture, so the last gate is whether the remaining 1.22× can be closed by lighting layout.

There are processes a strobe does not suit. In an indexing process where the part stops at the moment of capture, blur is absent in principle, so continuous light with a long exposure does better on both light quantity and stability, with no reason to push the chip with overdrive. In a configuration where low speed or coarse pixel resolution puts speed × exposure under 0.5 px, the trouble of adding sync wiring and a driver does not pay back. When the frame period gets very short and the duty cycle grows, average heat approaches the limit and the overdrive factor has to come down, so a line-scan structure that keeps the light on may do better. The overdrive factor needed varies with surface reflectance and the degree of LED efficiency loss, so the real light margin cannot be guaranteed before a sample test.

Tying speed, pulse width and pixel resolution into one line of multiplication is something you do several times a day in high-speed inspection, so I recommend getting this multiplication into your hands first. For whoever manages the cost, asking how many pixels the edge width grows at real line speed and whether there is a record of the measured light current waveform tells more about real detection power than the minimum exposure time of the camera. This article explains technology and is not investment advice. See our Disclaimer.

Field Note

To capture moving connector pin tips as if frozen, I would calculate the allowed blur and the pulse width before shortening the exposure and raising the gain. Even a 10 μs pulse may end up neither as bright nor as short as calculated if the rise and fall of the driver eat a large share of it. So the first step of setup is capturing the light current waveform to confirm the real emission time, and the threshold comes only after measuring that the calibration pattern edge width does not grow by more than 0.5 px on the move. After an aperture change, whether the pin height fits in the 4.5 mm depth gets checked again.

Field Checkpoints

Once pulse width, current waveform and depth of field are confirmed, I check the rest with the questions below.

• Is a WD of 110 mm or more secured by measurement, including the bar lights and the shading hood? (whether lights and hood fit between lens and part)
• Have the surface material and reflectance of the part been confirmed first? (glossy metal pins and matte solder mask need different light from the same pulse)
• Has the pulse width been back-calculated from the allowed blur (px) and travel speed and written in the spec? (12.25 μs or less in this setup)
• Have peak current, maximum pulse width and driver rise time been checked against the manufacturer rating together with the duty cycle? (not just average heat)
• Has the light current waveform actually been measured to confirm the effective pulse width? (the rising and falling parts are lost light and extra blur)
• After opening the aperture, does the depth of field still cover the pin and jig height variation? (4.5 mm at F/5.6)

Frequently Asked Questions

Can a strobe work with a rolling-shutter camera?

A rolling shutter starts exposure a little later row by row, so to freeze the whole image with one short flash, the pulse has to land in the interval when every row is exposing at once. Creating that interval means setting the exposure window longer than the row readout time, and ambient light builds up for that much longer, so shading matters more. Cameras differ in the modes that support this, so start with the trigger mode section of the spec sheet.

Does overdrive shorten LED life?

The SVS-Vistek notes also say lifespan may be affected and recommend keeping the on-time as short as possible. How much it shortens depends on the pulse rating of the manufacturer and the operating conditions, so there is no general answer. Instead, logging the mean brightness of a reference plate every frame lets a slow drop in brightness at the same current be caught early as a sign of chip degradation.

Glossary

• [1] Motion blur motion blur
  Smear in which the image drags along the travel direction because the target moves during exposure. Its length is speed multiplied by the time the light is on. ↩
• [2] Gain gain
  The factor by which the sensor signal is amplified electrically. It amplifies signal and noise together and only makes a dark image look brighter. ↩
• [3] Strobe strobe lighting
  A way of switching the light on once, very briefly, in step with the camera exposure. The time the light is on effectively becomes the shutter time. ↩
• [4] Overdrive overdrive
  A method that drives an LED briefly at a current above its rating to raise instantaneous brightness. It needs time to cool between pulses. ↩
• [5] Duty cycle duty cycle
  The share of total time spent switched on. Ten microseconds on, twenty times a second, gives 0.02%. ↩
• [6] Rise time rise time
  The time a driver takes to pull current up to its target value. The shorter the pulse, the larger the share this time takes. ↩
• [7] Depth of field depth of field
  The range of distance, in front of and behind focus, that still looks sharp. It narrows as the aperture opens. ↩
• [8] Global shutter global shutter
  A scheme in which every pixel of the sensor starts and ends exposure at the same instant. It suits capturing a moving target in a single moment. ↩
• [9] Trigger jitter trigger jitter
  Wobble in which the time between a trigger signal and the actual action varies slightly each time. On a moving target it turns into position error. ↩
• [10] WD working distance
  The distance from the front of the lens to the surface being inspected. Lights and the shading hood have to fit in between. ↩

References

• SVS-Vistek technical notes, Is there a reason for strobe controllers? (driving an LED above its rated current with short pulses, the effect on lifespan and the advice to keep on-time short, exposure sync and current control)
• Formulas: blur (px) = speed × pulse width ÷ pixel resolution, allowed pulse width = allowed blur × pixel resolution ÷ speed, duty cycle = pulse width × images per second, light energy ratio = (current factor × pulse width) ÷ (reference factor × reference exposure), light gain of one stop = (F1 ÷ F2)², depth of field ≈ 2 × f-number × circle of confusion × (1 + m) ÷ m², jitter position error = speed × jitter, magnification m = sensor width ÷ field of view, u = f × (1 + 1/m)

Related reading — Keeping Frame Rate and Lighting from Colliding on a High-Speed Inspection Line · On Thick Parts, Silhouette Measurement Error Comes from the Backlight’s Angle · On Moving Parts, Photometric Stereo Shrinks the Standstill Condition to 0.5 ms