Line scan camera resolution for micron-level semiconductor inspection

Lead frames, die attach and wafer scribes carry features a few microns wide. Miss a 12 μm bridge or a chipped pad corner and bad parts ship. Sizing that inspection is a pixel budget, not a guess: at least 3 pixels must span the smallest defect you have to catch, and everything else — magnification, image circle, line rate, exposure, camera bandwidth — follows from that number. This guide gives the rule, the formulas and a complete worked example: a 16K sensor resolving that defect across a 60 mm frame, with the lens, light and interface sized to match. A table shows where 8K and 12K sensors still hold, and the closing checklist covers the checks before you order.

The 3-pixel rule that sets the sensor

A defect exists in the image only if enough pixels land on it for a threshold to separate it from its surround. The working rule across machine vision is at least 3 pixels across the smallest feature you must detect; when the pixels must also produce a measurement, plan 5-8. Below roughly 2.5 pixels the feature sits mostly inside the lens edge spread. The line-to-line signal wobbles, and the same defect flickers in and out between scans. The general version of the rule, for area and line cameras at any scale, is worked in camera resolution for defect detection: size it by the pixel rule.

Three formulas turn the rule into hardware:

Ground sample distance (GSD) is the object width one pixel covers, in μm per pixel. Every downstream number — lens magnification, image circle, line rate, light power — hangs off it.

Worked example: 12 μm defects across a 60 mm frame

Take a lead frame strip 60 mm wide with a 12 μm defect specification: lead bridges, chipped pad corners, residue specks.

1. Required pixel pitch: 12 μm / 3 = 4 μm per pixel.
2. Pixels needed: 60 mm = 60,000 μm, and 60,000 / 4 = 15,000 pixels. The next standard line length up is 16K, 16,384 pixels.
3. Actual GSD: 60,000 / 16,384 = 3.66 μm per pixel, so the defect spans 3.3 pixels — the rule holds with margin.
4. Magnification: with the sensor pixel chosen below, m = 5 / 3.66 = 1.37×.
5. Sensor width: 16,384 × 5 μm = 81.9 mm, so the lens image circle must reach at least 82 mm.
6. Closing check: sensor width / m = 81.9 / 1.37 = 60 mm. The numbers agree.

The line scan field of view calculation runs step 2 in full for other strip widths, and the camera resolution calculator does the same arithmetic interactively:

The budget moves fast. Double the field to 120 mm and the pixel need doubles past any single line, so the strip splits across cameras. Halve the defect to 6 μm and the pitch budget halves to 2 μm, beyond the pixel pitch at any useful magnification — the same split, one level deeper.

The camera: 16K pixels, mono, F-mount

Step 2 asks for 15,000 pixels and the 16K class answers with 16,384. A 16K camera with a 5 μm pixel and an 82 mm image circle on an F-mount fits the worked example exactly.

The MV-CL166-91F2C carries that combination and outputs XoFLink on 2 SFP+. Mono is the right readout here: contrast on polished lead frames is reflectance, not colour, and mono puts every photon into one channel. The RGB/4TDI version earns its keep only when colour carries the defect signature or the extra TDI lines buy exposure. The camera finder filters line scan models by line length, pixel size and interface.

The lens: image circle, 100 lp/mm and 1.37×

Magnification 1.37× sets three lens requirements. Image circle first: the sensor is 81.9 mm across, so 82 mm is the floor, and every corner pixel lives near the edge of the optics where resolution falls first.

Second, the lens must resolve the pixel pitch at the sensor: 1 / (2 × 0.005 mm) = 100 line pairs per millimetre; a 3.5 μm pixel would ask 143 lp/mm, a 7 μm pixel 71 lp/mm.

Third, magnification fixes the geometry through f = WD × m / (1 + m). At 1.37× and a 173 mm object distance the focal length lands at 173 × 1.37 / 2.37 ≈ 100 mm, and the image distance sits near f × (1 + m) = 237 mm — about 410 mm front-to-sensor, the mechanical track the stage has to swallow (a thin-lens estimate; real line scan lenses focus internally). The lens focal length calculator reruns that step for your own working distance.

Lenses for 16K 5 μm sensors print their design numbers directly; this one carries 116 mm focal length, a 60-90 mm image circle and a 127.2 mm working distance on a V-mount:

Match the design magnification to your 1.37× within a few percent instead of stretching a lens past its design point. The line scan lens selection guide for 4K, 8K and 16K cameras compares the ranges by sensor class, and the line scan lenses category filters by mount, image circle and focal length.

Line rate, exposure and the light budget

On the scan axis, resolution comes from motion: line rate × GSD = stage speed. At a 50 kHz line rate and 3.66 μm per pixel the stage runs at 3.66 × 50,000 = 183 mm/s, exposure tops out at 1 / 50,000 = 20 μs, and motion blur is 183 mm/s × 20 μs = 3.7 μm — one pixel. The line rate set-up guide and the exposure time trade-off for moving parts carry the details; when an encoder drives the stage, encoder ticks per pixel fixes the scale.

Scan time closes: one 12 mm part is 12,000 / 3.66 = 3,279 lines, so 120 parts per frame is 393,480 lines, which is 7.9 s of scan at 50 kHz — inside a 25 s index-and-inspect cycle.

Twenty microseconds is a short exposure and the light has to earn it. High-intensity line lights exist for exactly this job: red LEDs convert drive power into light more efficiently than any other visible colour, which is why red dominates fast shiny-metal inspection. A lighting controller for line scan cameras strobes the LEDs to pack more photons into the 20 μs window. The power side of that trade-off is worked in how much power a fast line scan light needs. The light finder filters line lights by length, colour and voltage.

Keep the light longer than the field: a 60 mm FOV under a 300 mm line light leaves 120 mm of overhang each side, which is what stops the ends of each frame from dimming.

The LS-LSC-0527 prints 87 W of red or 108 W of white from a 300 × 17 mm emitting edge.

Bandwidth: where the interface decides

Pixels per line × lines per second is the data firehose. At 16K and 8-bit output a line is 16,384 bytes, so 50 kHz produces 16,384 × 50,000 = 819 MB/s, just inside the roughly 990 MB/s a 10GigE link moves after protocol overhead. Switch to 12-bit packed and the same scan runs 24.6 KB per line, 1,229 MB/s — past one 10GigE port. Doubling the line rate pushes even 8-bit out to 1.64 GB/s. That is what the 2 SFP+ XoFLink output and the CoaXPress options on line scan cameras are for; GigE Vision camera selection: bandwidth, cables and software plans the cable and the headroom. If the line you are upgrading already runs an older line scan camera, replacing a legacy line scan camera with a current model walks the same trade-offs as a retrofit.

The 8K path buys headroom: 8,192 pixels × 50,000 lines/s = 410 MB/s, half the 16K stream, comfortably inside 10GigE.

The LS-LCM-0099 is the 8K, 10GigE, mono, M72-mount representative of that class.

8K, 12K or 16K: the decision table

Run the same 60 mm field across the standard line lengths:

Column three is 60,000 μm divided by the pixel count; column five is column three × 3. Two escape hatches keep an 8K camera alive. Shrink the field to 30 mm — 30,000 / 8,192 = 3.66 μm, the identical GSD, with two cameras per lane — or accept a 22 μm defect floor. If the specification genuinely reads 12 μm, the table points at 16K; sensor selection for high-speed inspection adds the speed side, and the cameras topic page collects the sensor-choice guides.

Where the 3-pixel rule bends

Three honest exceptions, none of them free.

Sub-pixel edge fitting tracks an edge to about a tenth of a pixel, but it interpolates: detection still needs 2.5-3 real pixels of signal, so software does not buy back an undersampled sensor.

Binning trades resolution for signal: binning 2 × 2 halves the pixel count and roughly doubles the light per effective pixel. The pixel binning walk-through shows when that trade is worth taking.

TDI integrates one line across many sensor stages, multiplying effective exposure without motion blur — the standard answer when the line rate says 20 μs and the part cannot take more light. At low contrast it stacks noise too, so it helps most when the optics already deliver separation.

Checks before you order

A one-pass list that catches the expensive mistakes:

• Image circle at least the sensor width — 82 mm in the worked example — with a little margin, because corner resolution dies first.
• Lens Nyquist matched to the pitch: 100 lp/mm here, 143 lp/mm for a 3.5 μm pixel, 71 lp/mm for 7 μm.
• Light length at least the FOV, with overhang on both ends for uniformity.
• Exposure at or below the line period: 20 μs at 50 kHz keeps blur under one pixel.
• Data rate in your actual pixel format: 819 MB/s for 16K 8-bit at 50 kHz, 1,229 MB/s packed 12-bit — plan the interface for the second number if 12-bit is on the table.
• Encoder scale set so one tick equals one pixel or better, if the stage is encoder-driven.
• A trial scan of known-good and known-bad lead frames at production speed before the order closes.

The decision

For the defects in the worked example across a 60 mm field at production speed: a 16K mono camera at that pixel pitch; a lens with at least an 82 mm image circle near 1.37× — about 100 mm focal length at a 173 mm object distance; a high-intensity red line light longer than the field, strobed through a controller to fit a 20 μs exposure at 50 kHz; and a dual-port interface the moment the pixel format or the line rate doubles. A 12K sensor holds only to 15 μm features; an 8K holds only inside a 30 mm field.

Size the rest of the chain in the camera configurator, browse the line scan cameras and line scan lights categories, or send your field width and defect size through the contact page and we will run the numbers with you.