How to choose a camera for object detection

Object detection misses for two reasons: the object is too small in the image, or the sensor did not gather enough light to separate it from the background. Both are sensor properties, settled before any model runs. Most detection cameras in a given class share the same accelerator, so a TOPS figure cannot tell two models apart; the sensor format, the pixel pitch and the light on the part do. Check three numbers before you order: at least 3 pixels across the smallest feature (8-10 to classify), a pixel pitch of about 2.4 µm or more when light is short — a 2.4 µm pixel collects 2.9× the photons of a 1.4 µm pixel — and an exposure that keeps motion blur under half that feature (at 300 mm/s with a 2 mm feature, about 3.3 ms). This guide works the arithmetic through, then the smart-camera-versus-area scan camera decision and a short list of checks before you order.

The three numbers that decide detection

Judge a detection camera on three numbers, in this order. First, the smallest feature you must see spans at least 3 pixels across — 8 to 10 to classify it reliably. Second, the sensor collects enough photons per pixel to lift that feature above noise; this is pixel pitch and light, not resolution. Third, blur during the exposure stays under half the smallest feature. Everything else on the spec sheet — TOPS, frame rate, interface — only sets how fast and how easily you get that image.

A well-lit, static scene lets almost any modern camera from 1.6 MP up pass all three tests; choose on integration and cost. The choice narrows as light falls or speed rises — that is when pixel pitch, sensor format and shutter decide, and when the arithmetic below pays for itself. If you are building the station from scratch, the overview of what a machine vision system is and how to choose one places the camera among its lens, light and host.

Pixels across the object: how much resolution you need

The rule: 3 pixels across the smallest feature to detect it, 8-10 to classify or measure it. Worked example: a 300 mm wide field on a 4000-pixel sensor gives a pixel footprint of 300 mm / 4000 px = 75 µm. A 2 mm defect then spans 2 / 0.075 = 27 px — classification grade. Minimum horizontal resolution for detection alone: 3 px × (300 mm / 2 mm) = 450 px. Every camera class from 1.6 MP up clears that bar with room to spare.

So resolution is rarely the binding constraint in object detection; pitch and light are. Two cautions belong here. The lens must resolve the same detail — a 75 µm footprint asks about 1 / (2 × 0.075 mm) ≈ 6.7 lp/mm at the sensor, modest but not free. And more megapixels on the same format shrink the pixel, which is the subject of the next section.

Photons per pixel: the budget that beats TOPS

At a fixed scene brightness and f-number, the photons a pixel collects scale with its area — pitch squared. In the shot-noise-limited range where detection thresholds live, signal-to-noise ratio improves as the square root of the photon count:

Here N is photons per pixel, Φ the photon flux per square micrometre per millisecond, p the pitch in µm and t the exposure in ms. Take two typical pitches — 2.4 µm, a large-format industrial sensor, and 1.4 µm, a compact phone-class sensor — and an illustrative Φ = 2,000 photons/µm²/ms; scale linearly with your own light. Pixel areas: 2.4² = 5.76 µm², 1.4² = 1.96 µm², a ratio of 5.76 / 1.96 = 2.9×. At t = 1 ms: N = 2,000 × 5.76 = 11,520 photons against 2,000 × 1.96 = 3,920.

SNR = sqrt(11,520) ≈ 107 (40.6 dB) against sqrt(3,920) ≈ 63 (35.9 dB). To collect the same 11,520 photons, the small pixel needs 2.9 ms — 2.9× the exposure or 2.9× the light. That is the decision in dim or fast scenes: a 2.9× photon deficit is what flips a marginal-contrast feature between detected and missed from cycle to cycle. Choose a pitch of 2.4 µm or more when light is short; if a 1.4 µm-class pitch is fixed by the layout, budget triple the light or slow the line.

Hot and dim cells are the textbook case; GigE camera selection for high temperature and low light applies this same budget end to end.

Format and resolution set the pixel pitch

Listings print sensor format and megapixels; pitch often needs one division. The inch convention sets a sensor diagonal near 16 mm divided by the format denominator, a 4:3 sensor runs about 0.8 × diagonal wide, and pixels across = sqrt(megapixels × 4/3) × 1000. Example — 1/1.7" format, 12 MP: width = 12.8 mm / 1.7 = 7.53 mm; pixels across = sqrt(12 × 4/3) × 1000 = 4,000; pitch = 7.53 mm / 4,000 ≈ 1.9 µm.

The same method fills every row; a 3:2 sensor moves the estimate by 10-20%, and a sensor's own printed pitch wins where it exists.

Dense pitch is a deliberate choice, not a defect. The LS-CMR-0166 captures 4000 × 3000 at 30 fps over USB3 on a 1/1.7" sensor.

That is the 1.9 µm row: the right camera for a bright, static bench where 12 MP of detail buys detection margin and the host wants a simple USB3 link.

Smart camera or area scan camera

The second decision is where the model runs. A smart camera carries the accelerator on board and answers the PLC directly — no PC, no GPU licence, one device to mount. An area scan camera sends full frames to a host, which buys the widest sensor choice and the full C-mount lens range at the cost of a computer and integration work.

On-board compute is now effectively uniform across a smart-camera range — the same models run at the same rate on any of them — so pick on the sensor line: format, pitch, shutter, and whether integrated light suits the scene. The MV-SC3050XC-16M-ATN-H runs 5 MP colour on a 1/1.45" sensor with a global shutter and an M12 mount. That pairing works out near 3.4 µm per pixel — generous photons in a compact housing.

How to split the logic between a learned detector and classical checks is its own decision; the AI vs rule-based smart camera inspection comparison walks it.

When the photon budget or the pixel budget outgrows a compact format, move up. Three area-scan cameras that span this decision:

Read the trio left to right as one spread: a 2.5 MP colour sensor on a 1/1.7" format with a global shutter, a 6 MP mono sensor on 1/1.8", and 20 MP colour on a 1" format. From the pitch table: the 2.5 MP pairing lands near 4.1 µm — the low-light pick; the 6 MP mono pairing near 2.5 µm — the all-rounder; the 20 MP pairing near 2.5 µm on a much larger area — the detail pick when small features must be seen across a wide field.

A full build on this spread — camera, lens, light and a robot hand-off — is the screw thread inspection camera selection for robot handling application.

Exposure versus motion

Motion caps the exposure, and the photon budget must fit inside that cap. Keep blur under half the smallest feature for detection grade, or under one pixel footprint for measurement grade:

Worked example: belt at v = 300 mm/s, smallest feature 2 mm, pixel footprint 75 µm from the 300 mm / 4000 px field above. Detection grade: t_max = 0.5 × 2 mm / 300 mm/s ≈ 3.3 ms. Measurement grade: t_max = 0.075 mm / 300 mm/s = 0.25 ms. Now close the loop with the photon example: the 1.4 µm pixel needed 2.9 ms to match the 2.4 µm pixel at 1 ms — just inside the 3.3 ms window at this belt speed. Double the speed to 600 mm/s and the window halves to 1.7 ms; the small pixel can no longer reach the budget, so you add light, grow the pitch or accept the blur.

If the part moves during exposure, specify a global shutter: a rolling shutter reads rows sequentially and shears a moving part, which shifts and stretches the detection box. Rolling shutter is a static-bench choice. When you hand exposure to the camera's auto mode, know which pixels it averages — the auto exposure and auto gain on colour cameras: which pixels count explainer prevents surprises when a dark part enters a bright scene.

Mono or colour

Default to mono unless colour defines the class. A colour sensor's Bayer filter absorbs part of the spectrum before the photodiode, so the same silicon in mono collects roughly 1.5-2× more photons at visible wavelengths — the same square-root SNR gain as a pitch step, with no cost. Pick colour when the defect is a colour: missing print, mixed part, stain. Presence, orientation, geometry and count detect as well or better in mono.

Checks before you order

• Measure the field and the smallest feature, then compute pixels: 3 per feature to detect, 8-10 to classify.
• Derive the pitch from format and megapixels and hold it against the 2.4 µm line for dim scenes.
• Compute the exposure window from belt or robot speed, then check the photon budget fits inside it; if it does not, add light before you add camera — browse the machine vision lights range.
• Match shutter to motion: global for anything that moves, rolling for parts at rest.
• Check the interface against frame size: GigE moves about 118 MB/s, so a 12 MB mono frame means just under 10 fps; USB3 handles roughly 350-400 MB/s.
• Confirm mount and working distance: M12 for tight cells, C-mount for the full lens range.

When the obvious pick is wrong

More megapixels is not more detection. At the same 12 MP headline, a 1/1.7" format runs 1.9 µm pixels and a 1.1" format runs 3.5 µm — nearly four times the photons per pixel for the same resolution. When light is the constraint, format beats resolution.

The compact pick can beat the flagship. A smart camera with integrated light at short range often delivers more usable photons per pixel than a distant C-mount rig, because the light ships with the housing and the scene sits close and bright. And a presence check needs far less than you think: an object spanning 30 px detects fine at 1.6 MP with 3 µm pixels — no 20 MP sensor required.

The decision, per case

• Bright, static bench, features at 10 px or more: any modern camera from 1.6 MP up; choose on integration and cost.
• Dim cell, fast line, marginal contrast: pitch of 2.4 µm or more — 5 MP on 1/1.45", 12 MP on 1.1", 2.5 MP on 1/1.7" — mono preferred.
• Part or robot moving during exposure: global shutter, exposure window from the blur formula, light sized to fit it.
• No host allowed: smart camera with the largest format and pitch the budget allows.
• Over 12 MP, mono on a large format, or several cameras per decision point: area scan camera with a host or controller.

Match the rest of the chain on the camera finder, start a build in the camera configurator, or read the surrounding guides on the cameras topic page; the contact page is open when you want a second opinion on the numbers before you order.