Machine vision camera resolution: how many megapixels you need, and when more hurts.

How many megapixels do you need for machine vision? Divide your field of view by the smallest defect, multiply by about three pixels per defect, and do it for both axes. That pixel count, plus a little margin, is your camera. Most inspections land between 1.6 and 12 MP, and buying far more than the math asks for usually makes the station worse.

Our machine vision resolution calculator does that arithmetic for you, including focal length and exposure. This post covers the part a calculator can't. Which myths cost people money, why the rules of thumb are what they are, and how we'd size three real inspections. We're paid by no camera maker, so we've no reason to sell you pixels.

How many megapixels do I need for machine vision?

Take the field of view on each axis, divide by the smallest feature you must detect, and multiply by the pixels you'd want across that feature, usually three. A 120 mm field with a 0.2 mm defect needs 1,800 pixels across, so about 2.4 MP once you do the other axis. Then round up to a real sensor.

That three-pixel figure isn't ours. Teledyne DALSA's camera selection guide calls it a rule of thumb that many vision experts use, three pixels in the smallest area you want to detect.1 The reason is alignment. A defect exactly one pixel wide almost never sits neatly inside one pixel. It straddles two, each sees it at half contrast, and a bit of noise loses it. At three pixels, at least one is fully covered wherever the defect lands.

The number people get wrong is the field of view. They size it to the part, but parts arrive a few millimeters off nominal and the tool needs margin to find edges. On a lot of stations we've sized, position tolerance set the field more than the part did.

Fig. 01The resolution chain · the weakest link sets what you can see

Does a higher megapixel camera mean better accuracy?

Not by itself. More pixels over the same field shrink the millimeters per pixel, which helps detection and repeatability. Accuracy's a different thing. It depends on calibration, lens distortion, how flat and how repeatably the part sits, and whether the lens can actually resolve the smaller pixels.

Start with the lens, because it's the link people skip. A sensor can sample at most one line pair per two pixels, its Nyquist limit. For 3.45 µm pixels that's about 145 line pairs per millimeter; for 2.74 µm pixels it's about 182.2 A lens that delivers little contrast at that frequency turns your new pixels into a smoother picture of the same blur. Edmund Optics measured this on 2.2 µm pixels and found only 8.8 percent contrast at Nyquist, against 24 percent predicted from the lens's nominal MTF, and it recommends imaging at about half Nyquist for reliable machine vision.3 That's an optics company telling you not to count on your last pixel.

And a measurement's error budget rarely sits in the pixel count. A part that bows a millimeter, 2 percent lens distortion, a fixture that shifts as it warms up. Each can be bigger than a pixel, which is why we treat dimensional gauging with machine vision as its own discipline with calibration and a gage R&R, not as a megapixel purchase.

What pixel size should a machine vision camera have?

For general inspection, 3.45 µm is the safe middle. Pixels of 2.74 µm or smaller pack more resolution into a smaller, cheaper format, but they need sharper lenses, collect less light each and they'll hit diffraction sooner. Larger pixels, 4.5 µm and up, forgive lenses and low light at the cost of a bigger sensor.

Two Sony Pregius sensors make the trade concrete. The IMX264 and IMX547 both read out 2448 × 2048 pixels, so over the same field they give identical millimeters per pixel. The IMX264 has 3.45 µm pixels on a 2/3-inch format. The IMX547 has 2.74 µm pixels on 1/1.8-inch. Teledyne FLIR's EMVA 1288 measurements put their saturation capacity at 11,085 and 9,476 electrons.56 (EMVA 1288 is the standard way camera makers measure and publish sensor performance, so you can compare numbers across brands.4)

The full-well gap looks small. The bigger effect is light per pixel. At the same f-number and exposure, each pixel collects light in proportion to its area, and (2.74 ÷ 3.45)² is 0.63. Photon shot noise goes as the square root of the signal, so that pixel's signal-to-noise ratio drops to about 0.79 of the larger pixel's. To win it back you open the aperture and lose depth of field, lengthen the exposure and risk motion blur, or add light. None of that shows up on the spec sheet next to the megapixel number.

Fig. 02Pixel size vs light collected · same f-number and exposure

How does camera resolution affect frame rate and cycle time?

Roughly in proportion. Gigabit Ethernet carries about 125 MB per second (1,000 megabits divided by eight), and an 8-bit 5 MP image is about 5 MB, so you're capped at about 25 frames per second. A 24 MP image on the same link is about 25 MB, which leaves you under five frames per second.

Basler's own documentation shows it. The ace 2 a2A2448-23gmBAS, a 5 MP GigE camera, reaches 23.4 frames per second with its bandwidth settings tuned.7 Its 24.4 MP sibling, the a2A5328-4gcBAS, tops out at 4.8.8

Every frame has to be processed too, and inspection time grows with pixel count. A 24 MP image that takes 200 ms to arrive can take as long again to inspect, which is a whole cycle on a fast line. A region of interest or a faster interface helps, but if the arithmetic said 5 MP with margin, the cheapest fix is to buy 5 MP. It's also the one nobody suggests.

Fig. 03Frame rate vs megapixels on a GigE link

How many pixels per defect do you need?

About three for detecting a defect with decent contrast. One to two can work for a high-contrast presence check, five to ten for low-contrast defects or anything you need to classify, and gauging follows a different rule based on tolerance. It's contrast more than size that sets the right number.

These are our working numbers. If you're reading codes, our OCR and code reading guide goes deeper on module sizes. If you're weighing a neural network, deep learning vs rule-based vision covers how pixel needs change. And one honest warning. The table assumes the defect has contrast. Three pixels across a scratch that's almost the same gray as the metal still isn't a visible scratch.

Is lighting more important than camera resolution?

On most marginal inspections, yes. Resolution decides how many pixels land on a defect. Lighting decides whether those pixels are any different from their neighbors. A defect with high contrast survives a modest camera, and a defect with no contrast won't show up on any camera.

The usual story goes like this. A station misses scratches, someone decides the camera isn't sharp enough, and a bigger sensor goes in. It still misses scratches, because the scratches were lit straight on and looked like the rest of the brushed surface. Swing a low-angle bar light in and they jump out at the old resolution. Our machine vision lighting guide shows the geometry for each defect type, and the surface defect inspection page works through cosmetic jobs specifically. We'd spend the first dollar on light, the second on the lens and the third on pixels.

What camera resolution do I need for inspection? Three worked examples

Size from the defect up. Set the field from the part plus its position tolerance, pick pixels per defect from its contrast, then check frame rate, lens and motion blur before you buy. Here's that process on three common jobs. The parts and tolerances are illustrative, chosen to show the reasoning, not taken from a customer.

Example 1: cap presence and skew on a bottling line

Bottles pass at 600 a minute, 10 a second. A cocked cap's about 1 mm high on one side, and the field is the neck area, 60 × 45 mm with room for wobble.

At three pixels per millimeter of lift, you need 0.33 mm per pixel, or 180 pixels across. That's less than VGA. A 1.6 MP sensor at 1440 pixels across gives 0.042 mm per pixel, eight times what the job needs. At 80 mm bottle pitch the conveyor runs 800 mm/s, so keeping blur under one pixel at 0.042 mm per pixel means an exposure under about 52 µs. A 5 MP sensor over the same field pushes that to about 31 µs, which takes a lot more strobe. More pixels made this job harder. We'd buy a small global-shutter camera and put the money into a good backlight.

Example 2: scratches on a machined aluminum housing

Now a 120 × 90 mm face, with a cosmetic spec that rejects scratches wider than 0.1 mm and longer than 2 mm. Three pixels across 0.1 mm means 0.033 mm per pixel, and that's 3,600 × 2,700 pixels. That points at the 12 MP class, such as the Sony IMX304 at 4096 × 3000 with 3.45 µm pixels on a 1.1-inch format.9

Honestly, this one's a judgment call. Under low-angle dark-field light, a scratch scatters light into the camera and shows up bright even when it's narrower than a pixel, and it's 2 mm long, so it covers dozens of pixels along its length. A 5 MP sensor at 2448 pixels across gives 0.049 mm per pixel, two pixels across the width and about 41 along the length. We'd test both on real parts. If the 5 MP image shows the scratches clearly, it wins, because its 2/3-inch lens is cheaper and the camera runs faster. If they're faint, the 12 MP camera earns its keep.

Example 3: hole position on a sheet-metal bracket

Last, a 200 × 150 mm bracket with holes toleranced at ±0.1 mm true position. The 10:1 rule says the gauge should resolve a tenth of the 0.2 mm band, so 0.02 mm. With a backlight and clean edges, subpixel fitting to about a fifth of a pixel is realistic, so pixels of about 0.1 mm will do. A 5 MP sensor gives 0.082 mm per pixel across 200 mm, and a fifth of that is 0.016 mm. On paper, you're done.

But the bracket isn't flat. A standard lens at a 500 mm working distance changes magnification by about 0.2 percent for every millimeter the part moves toward it. If the bracket bows 1 mm, a hole 100 mm from the center appears to shift about 0.2 mm. That's the whole tolerance band, and no sensor fixes it. The answer is a fixture that clamps the bracket flat, or measuring in 3D. A telecentric lens would remove the effect, but its front element has to be as big as the 200 mm field, which gets expensive fast. Our guide on how to choose a machine vision lens covers when telecentric optics are worth it.

None of the three ended on the camera. That's typical. The pixel math gets you a shortlist, and the rest of the station decides whether it works.

When should you use multiple cameras or a line scan camera instead?

Use more than one camera, or move one camera, when a single sensor can't put enough pixels on the defect across the whole field. Use a line scan camera when the part moves continuously past the camera, is very long, or can be rotated, like a web, a sheet or a cylinder.

Say you need 0.12 mm pixels across a 1,000 mm wide web. That's about 8,200 pixels across, and an area camera that wide is a big, slow, expensive sensor that also needs a big, even patch of light. A line scan camera like Teledyne's Linea 8k has 8,192 pixels in one row and runs up to 13 kHz on GigE.10 Over 1,000 mm that's 0.122 mm per pixel, and for square pixels at 1 m/s you need about 8,200 lines a second, comfortably inside its rating. The light becomes one bright line instead of a square meter, though you need an encoder to tie line rate to speed.

For a part that sits still, two cameras or one moved camera often beat one giant sensor. Each view keeps a normal lens and normal light. In a robot cell it's often cheaper to put one 5 MP camera on the arm and take three pictures, which is how we build a lot of vision-guided robotics stations. The smart camera vs PC-based vision comparison explains why the PC route gets cheaper per added view, and 2D vs 3D machine vision covers jobs where the missing dimension is height, not pixels.

Fig. 04One camera, several cameras or line scan

How do you choose a machine vision camera without overbuying?

Write down the smallest defect, its contrast, the field including part position tolerance, and the cycle time. Size pixels from those, pick the largest pixel size that fits a common lens format, then test the camera on a stand with real good and bad parts before you order. Don't skip that.

That stand test is the cheapest step in the project. Twenty minutes with the real light at the real working distance catches contrast and depth-of-field problems no spreadsheet will, and it settles arguments like Example 2 with evidence.

Be wary of a quote that leads with megapixels. Ask the vendor to show you the pixels-per-defect math. If you're comparing systems, our pages on the best machine vision systems for inspection and Cognex vs Keyence cover what each platform offers at each resolution, and the machine vision system cost guide shows where pixels sit in the total bill. An oversized sensor running at the edge of its frame rate is also one way stations end up with loosened thresholds and false rejects of good parts.

In a robot cell we size the camera with the cell, because the robot can present the part or move the camera, and that's a different field. Our machine vision capability page shows how we build that in, and the machine vision library has sizing worked through for each inspection type.

Frequently asked questions

What is the difference between megapixels and spatial resolution?

Megapixels count the pixels on the sensor, such as 2448 × 2048 for a 5 MP camera. Spatial resolution is the size of one pixel projected onto the part, in millimeters per pixel, and it's what decides whether a defect is visible. The same 5 MP camera gives about 0.05 mm per pixel over a 120 mm field and about 0.008 mm per pixel over a 20 mm field.

Is a color camera lower resolution than a monochrome camera?

Effectively, yes. Most color cameras put a Bayer filter over the sensor, so each pixel sees only red, green or blue and the missing colors are interpolated from its neighbors. Fine detail's softer than on a monochrome camera with the same pixel count, and each pixel gets less light. Use monochrome for gauging and fine defects, and color only when the decision depends on color.

Can a region of interest make a high-resolution camera faster?

Usually, yes. Most industrial cameras can read out just a region of interest, and the frame rate rises as fewer rows cross the interface. It's a good fix when the defect can only be in a strip of the full image. It won't help if every frame needs the whole field, and processing time still scales with the pixels you keep.

What is subpixel accuracy in machine vision?

It's locating an edge or a pattern to a fraction of a pixel by fitting the gray-level transition across several pixels. On a clean, backlit edge, repeatability of a tenth to a fifth of a pixel is a realistic target. It improves measurement, not detection, and it depends on contrast, calibration and a part that sits still, so prove it with a gage R&R.

Do I need a global shutter camera for inspection?

If the part or camera moves during the exposure, you'll usually want one. A rolling shutter exposes rows one after another, so a moving part comes out skewed. A short strobe in a dark enclosure can freeze motion even with a rolling shutter, but a global shutter removes the problem. For parts that sit still, a rolling-shutter sensor is often a cheaper way to get high resolution.

What resolutions do machine vision cameras come in?

Industrial area cameras cluster around a few sensor classes. Common ones are 1.6 MP (1440 × 1080), 5 MP (2448 × 2048), 12 MP (4096 × 3000), 20 MP (5472 × 3648) and about 24.5 MP (5328 × 4608), mostly built on Sony sensors. Line scan cameras aren't sized that way. They're specified by line length, typically 2k, 4k, 8k or 16k pixels. Our resolution calculator lists the common formats.