Spatial Resolution in Machine Vision: mm per Pixel, Minimum Detectable Size, and the Diffraction Limit

How do you calculate mm per pixel?

Divide the field of view by the pixel count along the same axis: mm per pixel = FOV (mm) ÷ pixels. A 100mm wide scene on 2592 horizontal pixels gives 0.0386mm per pixel. The reciprocal, pixels per mm, tells you how many pixels cover each millimeter of scene.

Use mm per pixel for inspection sizing. It answers "if I see a bright region three pixels wide, how big is the defect?" Use pixels per mm when comparing against lens specifications, because lens resolving power is quoted in line pairs per millimeter (lp/mm) and one line pair needs two pixels (one dark, one light) to be sampled.

Model mm per pixel for any sensor, focal length, and working distance with the Commonlands FOV calculator, and see the field of view guide for how active sensor dimensions determine real scene coverage. For lenses with noticeable distortion, the average mm per pixel does not hold at the frame edges: local magnification varies across the field.

What is angular resolution (IFOV) in machine vision?

Angular resolution, also called instantaneous field of view (IFOV), is the angle subtended by a single pixel: IFOV = FOV ÷ pixel count. An 80° horizontal FOV across 3840 pixels gives 0.021° per pixel, or 48 pixels per degree. Multiplied by the working distance, IFOV becomes the pixel footprint on the object.

IFOV and its inverse, pixels per degree, describe the same angular sampling density. Pixels per degree is often the more intuitive form for detection work because it answers "how many pixels land on a target that spans X degrees?"

The telephoto gives ten times the angular detail of the fisheye over one tenth the scene width. At fixed sensor resolution, wide coverage and fine angular sampling are in direct tension. The task requirements decide which wins.

From IFOV to pixels on target

IFOV connects directly to the pixel-coverage rules above. A target's angular size is roughly 57.3 × feature (mm) ÷ WD (mm) in degrees; multiply by pixels per degree to get pixels on target. A 10mm feature at 500mm subtends about 1.15°, so a camera delivering 48 pixels per degree puts roughly 55 pixels on it, a comfortable margin over the 2–3 pixel floor for basic detection. Because contrast, noise, and blur are worse in production than on the bench, it is good practice to design for roughly 2× to 3× the theoretical minimum pixel count rather than the bare minimum.

IFOV = FOV ÷ pixels is an average across the sensor, and the local value varies by projection. Ideal f-theta and equidistant lenses map angle linearly to image height (r = f × θ), so degrees-per-pixel stays constant from center to edge. That uniformity is the defining property of the f-theta mapping. Rectilinear lenses (r = f × tan θ) devote more image height per degree toward the edge, so edge IFOV is smaller than center IFOV: at 40° off-axis, dr/dθ = f × sec²(θ) compresses the edge IFOV by about 1.70× versus center. Equisolid lenses (r = 2f × sin(θ/2)) move the opposite direction, growing edge IFOV. Real fisheye lenses deviate from any ideal mapping, so a detection task that gets 10 pixels on target at frame center may get more or fewer at the corner depending on the specific lens design, not by defect. The wide-angle and fisheye distortion guide covers projection models in detail; characterize local IFOV from the measured distortion profile for any lens with heavy distortion.

Angular resolution is also the more stable specification when working distance varies: mobile robots, drones, and vehicle cameras rarely image at one fixed distance. See lenses for robotics for how detection range drives lens selection in those systems.

What does a megapixel rating mean on a machine vision lens?

A megapixel rating on a lens names the sensor class it was designed to cover (a pixel pitch plus a format), not its optical resolving power. Resolving power is measured in line pairs per millimeter at the image plane, and the Nyquist rule sets the requirement: lp/mm = 1 ÷ (2 × pixel pitch in mm).

Sony's Pregius S family shows why the label misleads. The IMX547 (5.1MP), IMX546 (8.1MP), and IMX541 (20.4MP) all use 2.74µm pixels, so all three need a lens resolving at least 182 lp/mm. What differs is the image circle: the 20.4MP sensor is physically larger and needs wider coverage. A lens built for the IMX547 mounted on the IMX541 produces dark corners, not from lack of resolving power but from an image circle that is too small. Same optics, different megapixel label.

The second, independent requirement is coverage. The lens image circle must be at least as large as the sensor diagonal, or the corners vignette. Format names are carryovers from vidicon tube cameras and do not equal the diagonal in millimeters:

Matching a lens to a sensor: two checks, both required

1. Find the pixel pitch. Look in the sensor datasheet, not on the camera marketing page. Common values are 3.45µm (Pregius), 2.74µm (Pregius S), and 2.0µm (embedded 8MP class).
2. Calculate the required lp/mm. lp/mm = 1 ÷ (2 × pitch in mm). Treat the result as a minimum and prefer lenses with published margin above it.
3. Find the sensor diagonal. Take it from the active-area dimensions in the datasheet, or from the format table above. The sensor size and lens compatibility guide maps common sensors to required coverage.
4. Verify both on the lens datasheet. Resolving power must be ≥ the required lp/mm and the image circle ≥ the sensor diagonal. If either fails, the lens limits the system regardless of the megapixel labels involved.

Why lens MTF, focus, and lighting set the real limit

Pixel math sets the ceiling on system resolution. Lens MTF, focus accuracy, illumination contrast, and distortion decide how close a real system gets to it, and each one raises the practical minimum detectable size above the formula result.

MTF describes the contrast a lens produces as a function of spatial frequency. If the lens holds only 5% contrast at the sensor's Nyquist frequency, those pixels record a nearly flat signal; adding more pixels pushes Nyquist higher and captures more noise, not more scene. MTF also varies across the field (a lens can deliver 60% contrast at 145 lp/mm in the center and 20% at the corner), so verify performance at the image positions where defects actually appear. The MTF curve guide covers how to read those datasheets.

Defocus spreads each scene point across multiple pixels. A blur circle two pixels wide roughly halves effective spatial resolution, and scenes with depth make this a standing constraint through the circle of confusion. The two lens families handle it differently: C-mount lenses refocus through an internal cam system that moves lens groups relative to each other, and many offer an adjustable iris for depth-of-field control; M12 lenses are rigid assemblies focused by threading the whole lens in its holder, with a fixed aperture chosen at purchase. Neither approach removes the tradeoff. It changes which adjustments are available in the field.

Illumination determines the contrast a defect presents before the lens ever sees it. A 0.1mm scratch invisible under diffuse bright-field lighting can be plainly visible under dark-field grazing illumination with the same lens and sensor. Backlighting, structured light, and polarization each recover contrast that pixel coverage alone cannot. Design lighting and spatial resolution together, not sequentially.

Distortion makes spatial resolution position-dependent. Local magnification at the corners of a wide-angle frame can differ meaningfully from the center, so mm per pixel calculated as an average may be optimistic exactly where the feature sits. For measurement tasks, check resolution in the actual region of interest. A low distortion lens keeps the mapping close to uniform and simplifies calibration.

Common mistakes when specifying resolution

A few specification errors show up again and again. Each traces back to treating one layer of the resolution chain as the whole answer.

• Specifying megapixels without a field of view. Megapixels are a pixel budget, not a distribution. A 20MP camera behind a wide lens can deliver coarser mm per pixel at the part than a 5MP camera behind a telephoto. State pixel count and FOV together.
• Treating 2 pixels as a detection threshold. Two pixels is the Nyquist floor, reachable only with near-perfect contrast and grid alignment. Design for 3 to 5 pixels minimum, and more for low-contrast defects.
• Matching lens to sensor by megapixel label instead of pixel pitch and image circle. The label bundles two independent properties. Verify lp/mm against the Nyquist requirement and image circle against the sensor diagonal.
• Stopping down for depth of field without checking the Airy disk. Past roughly two pixel pitches of diffraction blur, each additional stop trades away fine-detail contrast the algorithm needs. Compute d = 2.44 × λ × F/# before committing to F/11 or F/16.
• Assuming uniform resolution across the frame. Corner MTF is lower than center MTF on many lenses, and distortion shifts local magnification. Verify performance at the image positions where defects actually appear.
• Ignoring motion blur in the exposure budget. A system sized for a stationary target loses its margin when line-speed smear exceeds half a pixel. Check exposure time against velocity before finalizing the illumination spec.
• Solving for detection when the task is measurement. Detection succeeds at 3 to 5 pixels; dimensional measurement typically needs 10 or more plus sub-pixel processing. Confirm which deliverable the downstream process expects.