Quad Bayer 4-cell sensors: binning, remosaic and full resolution

Quad Bayer — the layout most data sheets call 4-cell — puts four same-color pixels in every 2×2 cluster of a 4×4 cell. Buyers keep asking whether that forces 2×2 binning in silicon and gives up resolution for light. It forces neither: the filter layout and the readout mode are separate design choices, and a 4-cell sensor is built to run binned or at full resolution. This explainer separates the two, puts numbers on what binning buys (4× the signal, a quarter of the pixels) and what remosaic costs (interpolated colors), works a 12 MP example, and lists the mode checks to run before you commit a design.

Does a 4-cell CFA mean the sensor always bins 2x2?

No — the color filter layout and the readout mode are independent circuits, and 4-cell sensors are built to run both ways: 2×2 binning at quarter resolution with 4× the signal, or full-resolution readout with a remosaic step in front of the demosaic.

The color filter array is a layer of dyed filters over the photosites. The readout mode is a sequencer setting: which rows and columns the ADC chain digitizes, and what gets summed on the way out. Nothing in the pattern forces the sum. What decides your options is the register set — the output-format table that lists every supported resolution and the binning behind it.

What the 4-cell layout changes is which sums make sense in color. A 2×2 block of a standard Bayer grid holds one red, one blue and two green photosites; summing it mixes three colors into gray. Summing same-color neighbors there means combining pixels two pitches apart on the diagonal. In a 4-cell pattern each 2×2 cluster is one color, so the hardware sums a cluster and hands you an ordinary RGGB macropixel with four times the charge. That is why native color binning arrived with 4-cell silicon — not why you must use it.

Demand the output-format table for any part you shortlist. If both the full-resolution mode and the binned mode are listed, you get a two-mode camera behind one optic; if only one is listed, that part runs one way, and you decide now which way you need.

How is a Quad Bayer layout different from standard Bayer?

A standard Bayer mosaic repeats one 2×2 cell — two green, one red, one blue — while a 4-cell pattern scales that same cell to 4×4, giving each color a 2×2 cluster of its own; both still spend 50% of the photosites on green.

The classic Bayer mosaic tiles its 2×2 cell across the whole array. Scale that cell two times in each axis and replace every single filter with a same-color 2×2 cluster: you have the 4×4 quad cell, 16 photosites, four per color. Nearest same-color neighbors now sit one pitch apart instead of two, which cuts optical and electrical crosstalk and lets a cluster share microlens and conversion-gain tricks that lift low-light response. Published quad-CFA patent filings also drive those clusters for phase-detection autofocus and per-pixel aperture control — the features that sold the layout to the phone market.

How does the sensor deliver full resolution from a quad pattern?

It reads every photosite at native count — 4000 × 3000 stays 4000 × 3000 — and a remosaic stage re-sorts the quad grid into a conventional Bayer grid before ordinary demosaicing; nothing is binned and no photosite is dropped.

Where the remosaic runs is a per-camera fact: on the sensor's ISP, in camera firmware, or in your host library. In every case the arithmetic is the same — the samples at standard Bayer positions are interpolated from the measured quad samples. Luminance survives well because every photosite contributes; chromatic detail at high spatial frequency is synthesized. Frame rate pays for the pixel count: full-resolution readout moves four times the data of binned readout, so expect roughly a quarter of the binned frame rate on the same chain.

Full-resolution measurement does not require a quad part at all. The mono MV-CS200-10UM reads a 20 MP frame from a 1" sensor with no color interpolation in the chain, and the 12 MP LS-CMR-0166 delivers the same 4000 × 3000 geometry over USB3.

Both hold their rated output on every frame.

How much sensitivity does 2x2 binning really add?

Summing the four same-color pixels gives 4× the collected charge — two stops of exposure — and a 2× signal-to-noise ratio (+6 dB) in shot-noise-limited light, because noise grows only as the square root of the signal.

An exposure of 20 ms at full resolution holds its signal-to-noise ratio at 5 ms binned, so a 50 fps framing limit becomes 200 fps whenever the readout chain keeps up. The summed full well is four times larger too, so highlight headroom grows by about two stops before clipping. In dim, short-exposure work where read noise dominates, charge-domain binning converts one cluster instead of digitizing four pixels, and the gain approaches 4×.

Bandwidth follows the same ratio: a binned frame is a quarter of the bytes, which is what makes binning attractive on a crowded link. For how that trade-off compares with ROI and subsampling, read ROI vs binning vs subsampling: choosing faster acquisition.

What does binning cost in measurement resolution?

Half the sampling density in each axis: a 4000 × 3000 sensor outputs 2000 × 1500, so the same field of view goes from 12 μm per pixel to 24 μm per pixel, and the smallest measurable feature roughly doubles.

Worked example. Take a representative quad sensor with 4000 × 3000 photosites of 1.8 μm pitch viewing a 48.0 mm field of view:

• Full resolution: 48.0 mm ÷ 4000 px = 12 μm per pixel.
• Binned 2×2: 48.0 mm ÷ 2000 px = 24 μm per pixel, from 3.6 μm macropixels.
• A 0.10 mm pore spans 8.3 px full resolution — measurable — but 4.2 px binned — detect only.
• A 0.25 mm feature spans 10.4 px binned, enough to measure.

The decision for this line: a spec of "find 0.10 mm defects" runs full-resolution remosaic; a spec of "confirm a 0.25 mm feature" takes the binned mode and its four times the light. The arithmetic is identical for any 12 MP frame, for example the 4000 × 3000 output of the LS-CMR-0166 at 30 fps over USB3.

Run your own field of view and defect size through the pixel calculator, then set the mode to match the spec.

Small pixels promise more detail than they always deliver; small pixel camera selection: resolution versus usable detail covers where that gap bites.

Do remosaic and demosaic artifacts affect measurements?

They do at color boundaries and fine periodic detail: remosaic synthesizes the missing color samples, so a monochrome edge can pick up 1-to-2-pixel color fringes, and periodic patterns near the sampling limit can alias into false colors or moiré.

Every color camera interpolates — a Bayer sensor measures one color per photosite and computes the other two. Quad adds one interpolation earlier in the chain, so high-frequency chroma passes through two synthesized stages. Keep pass/fail decisions off computed color values on features two pixels wide, light the part with broadband white so all three channels see the edge, and treat mesh, thread and halftone textures as the worst case: their periodic structure aliases first, and quad's wider color spacing lowers the frequency where it starts. Color camera spatial resolution: beyond the pixel count quantifies what a CFA really costs in line pairs; if you take raw output, debug packed Bayer acquisition before image processing covers the format surprises on the way in.

When should a machine-vision system use a 4-cell sensor?

Choose quad when one camera must do two jobs — a low-light 3 MP overview and, on demand, a 12 MP detail view through the same optic — and skip it when the defect spec lives at the fine end of the full-resolution mode.

• One optic, two modes: quad. Binned frames for presence and counting, full-resolution frames when something needs a closer look.
• Light you cannot add — night work, unlit corridors, photons limited by the part itself: quad binned buys two stops with no new hardware.
• Light you control: skip quad. A strobe at full output beats two stops of binning and keeps full sampling density.
• Moving parts: skip quad. 4-cell parts are rolling shutter, and skew on a moving web distorts geometry; a global-shutter mono sensor measures straight.
• Depth or 3D pipelines: recalibrate per mode. Binning changes the effective pixel pitch, so intrinsics captured at full resolution will not reconstruct depth from binned frames — camera intrinsics after binning: reconstruct depth correctly shows the correction.

Compact color modules — board-level cameras built around phone-class silicon — are where 4-cell usually reaches vision hardware. The LS-BCM-0056 module carries 12 MP with 1.6 μm × 1.6 μm pixels behind a C-mount.

What should you verify in the readout mode list before buying?

Get the output-mode table — full-resolution remosaic, 2×2 binning, any skip modes, with the frame rate of each — and confirm where the remosaic runs, because that register list, not the CFA geometry, is what your software will see.

1. Every output resolution with its frame rate, in writing, before layout.
2. Where remosaic runs: on-sensor ISP, camera firmware, or your host library — and whether raw quad output is available at all.
3. Shutter behavior at your line speed, judged from sample frames, not from the pattern diagram.
4. A slanted-edge and a zone-plate target shot in both modes; fringes and moiré show up in minutes.
5. The mode listed in the camera's GenICam or USB3 Vision format list, not only in the sensor's NDA paperwork.

Phone-class sensor documentation sits behind NDAs — the roadblock most small teams hit first. The practical route is to buy the die inside a camera with an open vision interface: the camera maker has already done the NDA work, and the mode list appears in the documented feature set you can read from your own code.

Compare candidates across area scan cameras in the camera finder, or start at the camera step of the system configurator with both modes in mind. For a mode question on a specific part, contact the applications desk.