Understanding resolution and field of view

Resolution and field of view are two different things

In machine vision, "resolution" is overloaded, and confusing its meanings is the most common sizing mistake. Sensor resolution is simply the pixel count (for example 5 MP, or 2448 x 2048 px). Field of view (FOV) is the physical extent of the scene the camera images, expressed in millimetres. Neither tells you, on its own, whether you can see the defect that matters.

The quantity that actually decides detection capability is spatial resolution: how many millimetres each pixel covers on the part. You get it by dividing the FOV by the pixel count along the same axis. A 2448 px wide sensor looking at a 245 mm wide field gives roughly 0.1 mm/px. That single number — not the megapixel figure — is what you compare against the size of the feature you need to find.

Always size from the application inwards: define the part, the FOV that guarantees the part fits with tolerance for handling and positioning, and the smallest feature you must resolve. The sensor falls out of those constraints, rather than being chosen first.

Sizing the sensor: the 3-5 pixel rule

A feature one pixel wide is invisible in practice — sampling, optical blur and noise will swallow it. As a robust engineering rule, allocate at least 3 to 5 pixels across the smallest feature or defect you must detect reliably. Use 3 px for a forgiving, high-contrast feature and 5 px (or more) when contrast is low, the defect is subtle, or a measurement must be accurate and repeatable.

The arithmetic is straightforward. Required pixels along an axis = (FOV along that axis / smallest feature size) x pixels-per-feature. Round up to the next standard sensor and repeat for both axes; the larger requirement sets the sensor. The example below shows the full chain from part to part number territory.

• Define FOV with margin: largest part dimension plus 10-20% for part-to-part variation and mechanical placement.
• Choose pixels-per-feature: 3 px for easy detection, 4-5 px for low contrast or for gauging/measurement tasks.
• Compute needed pixels per axis: (FOV / smallest feature) x pixels-per-feature.
• Pick the next standard resolution above both axis requirements; confirm the resulting mm/px meets your accuracy budget.
• Worked example: 100 mm FOV, 0.2 mm smallest defect, 4 px/feature -> (100 / 0.2) x 4 = 2000 px on that axis.

Why optics and lighting cap usable resolution

Pixels only count if the lens can deliver detail to them. A lens has its own resolving power, set by its design, aperture and the contrast it preserves at fine detail (its MTF). Pairing a high-megapixel sensor with a low-resolution or undersized lens simply produces a sharp-looking image that is actually blurred at the pixel level — you pay for pixels you can never use. The lens must be rated for the sensor's format and pixel size; a lens that is fine on a coarse-pixel 2 MP sensor may be the limiting factor on a fine-pixel 12 MP sensor of the same format.

Lens choice also governs geometry. For dimensional measurement, a telecentric lens removes perspective error and the magnification change with working distance that a standard C-mount entocentric lens introduces, so the mm/px stays constant across the depth of the part. That is often what makes a stated spatial resolution achievable in practice rather than just on paper.

Lighting is the other hard ceiling. Resolution is meaningless without contrast: a defect that is the same brightness as its background cannot be resolved at any pixel count. Correct geometry — directional, diffuse, dark-field, dome, backlight or coaxial — creates the contrast that turns pixels into detections. Stable, sufficient illumination also lets you keep exposure short and the aperture stopped down, which preserves both the lens MTF and the depth of field. Specialist lighting brands such as DCM Sistemes exist precisely because this stage, not the sensor, is frequently where usable resolution is won or lost.

Megapixels vs sensor format: the trade-offs

Two sensors can share a megapixel count yet behave very differently because of pixel size and optical format. Packing more pixels into the same format shrinks each pixel, which reduces the light gathered per pixel (lower sensitivity, more noise in dim conditions) and demands a sharper, often more expensive lens to feed them. Spreading the same pixel count over a larger format keeps pixels big and tolerant, but needs a larger-format lens and can reduce depth of field.

Format also dictates lens availability and cost. Common C-mount lenses comfortably cover up to roughly 1-inch / Type 1.1 sensors; larger formats push you toward F-mount or dedicated large-format optics that cost significantly more. It is good practice to confirm the lens covers the chosen sensor format before committing to a high-resolution body, because a high-MP sensor that out-resolves its lens or vignettes at the corners is wasted capability.

• Smaller pixels: more resolution per format, but lower sensitivity, more noise and stricter lens demands.
• Larger pixels / larger format: better light performance and dynamic range, but pricier optics and shallower depth of field.
• Match the lens image circle and resolving power to the sensor format and pixel pitch — not just to the mount type.
• For high pixel counts at speed, weigh global vs rolling shutter and the interface bandwidth together.

Why more megapixels is not always better

Beyond optics, every extra pixel costs bandwidth, processing and money downstream. Data rate scales with resolution times frame rate times bit depth, and it quickly outgrows the interface. As a rough guide, GigE Vision carries about 1 Gbit/s over up to ~100 m with PoE; USB3 Vision offers roughly 350-400 MB/s over short 3-5 m cabling; and high-resolution-at-high-speed combinations push you to CoaXPress (CXP-6/CXP-12, multiple GB/s) with a frame grabber. Choosing a sensor without sizing the link is how projects discover, late, that they cannot achieve the required frame rate.

Larger images also raise CPU/GPU load, memory and cycle time. A vision system that must keep up with line speed has a fixed time budget per part; doubling the pixels can double transfer and processing time and break that budget, even though the image "looks better". For deep-learning inspection in particular, oversized images inflate inference time without improving results once the defect is already well sampled.

The disciplined approach is to resolve the smallest feature with 3-5 px and stop there. Spend the saved budget on better optics, lighting and a faster interface rather than on pixels you will throttle, crop or downsample anyway. If you are unsure how the camera, lens, lighting and interface interact for your part, AIS Vision can validate the full imaging chain, and where AI classification is the right tool, Rosepetal handles the defect-detection layer once the optics are sound.