High-resolution imaging for inspection

When 20+ MP actually makes sense

High resolution is a means to an end, not a goal. The right question is never "how many megapixels?" but "what is the smallest feature I must reliably detect, and over how large an area must I see it in a single image?" When those two requirements collide - a sub-millimetre defect spread across a panel, wafer, web or large assembly that you cannot break into multiple smaller views - a 20, 25, 50 MP or larger sensor becomes the economical answer.

Typical drivers are large-area inspection where moving the part or the camera is impractical, metrology that needs many pixels across a wide measurement range, and applications such as flat-panel display, PCB, semiconductor, solar and printed-web inspection. If your field of view is small or your part is presented in discrete poses, a standard 2-12 MP camera will usually do the job at far lower cost and complexity, so reserve high resolution for cases where the field of view and feature size genuinely demand it.

Sizing the sensor from the defect, not the spec sheet

Start from the smallest defect or feature you must resolve and apply the 3-5 pixel rule: a feature should span roughly 3-5 pixels to be detected and measured robustly rather than aliasing into a single ambiguous pixel. Divide your field of view by the smallest feature size, multiply by that pixel factor in each axis, and you have the pixel count the application demands. A 300 mm field that must resolve a 50 micron defect at 4 px, for example, points to roughly 24,000 px across - well beyond any single area-scan sensor, which is exactly when line-scan or multi-camera tiling enters the conversation.

Only once you know the required pixel count should you look at sensors. Large-format CMOS sensors (Sony Pregius S, Gpixel and similar families) reach 20-25 MP and beyond, with the very high counts moving to large optical formats. Remember that more, smaller pixels are not free: smaller pixels collect less light and can have lower full-well capacity, so very dense sensors may need brighter, more uniform lighting and longer exposures to maintain signal-to-noise. Match pixel size and sensor format to the optics and lighting you can realistically deploy.

• Smallest feature to resolve (in object units) - the hard requirement.
• Field of view in both axes - sets the magnification and total pixels.
• Pixels per feature - use 3-5 px for detection/measurement reliability.
• Resulting pixel count - rounds you up to a candidate sensor format.
• Pixel size and full-well - drive light budget and noise; check at your target exposure.

The optics challenge: resolving power, MTF and image circle

A high-megapixel sensor only delivers if the lens can resolve detail at the sensor's pixel pitch. This is the most common high-resolution mistake: pairing a 25 MP sensor with a lens that cannot project that much detail, so the image is large but soft and the extra pixels carry no new information. Specify lenses by their resolving power and MTF at the relevant spatial frequency (line pairs per millimetre that correspond to your pixel pitch), not by focal length alone, and prefer lenses explicitly rated for the sensor's megapixel class.

The second optics constraint is the image circle. A large sensor needs a lens whose image circle covers the full diagonal without vignetting or corner blur; a standard C-mount lens designed for small sensors will darken and soften the corners of a large format. Larger sensors frequently require larger mounts (such as F-mount, M42, M58 or TFL) and purpose-built large-format optics. Where measurement accuracy and edge geometry matter, a telecentric lens removes perspective error and keeps magnification constant with working distance - at the cost of an entrance aperture as large as the field, which becomes expensive at high resolution. Plan optics and budget together: on a high-resolution system the lens can rival or exceed the camera in cost.

Bandwidth and interface: feeding the data firehose

A 25 MP image is roughly 25 MB at 8-bit mono, more at 10/12-bit or colour, so even modest frame rates produce enormous data streams. A single GigE Vision link (~1 Gbit/s, around 100-120 MB/s usable) is quickly saturated - a 25 MP mono camera maxes out at only a few frames per second over GigE. USB3 Vision (~350-400 MB/s) buys more headroom but is limited to short 3-5 m cable runs and one camera per host port. For meaningful frame rates at high resolution you generally move up to 10GigE, or to CoaXPress.

CoaXPress (CXP-6, CXP-12 and multi-link configurations) delivers multiple GB/s over coax with power and trigger on the same cable, but it requires a frame grabber in the host PC - a real cost and integration factor, not an afterthought. Choose the interface from the data rate your resolution and frame rate demand, then confirm the host can actually consume it: PCIe lanes, NIC or grabber capacity, CPU/GPU for any on-the-fly processing, and disk throughput if you archive images. Brands such as Basler, iRAYPLE and Omron / Sentech offer 10GigE and CXP-12 cameras in the high-resolution class, paired with matching frame grabbers.

• GigE Vision: ~1 Gbit/s, up to ~100 m, PoE - only a few fps at 20+ MP.
• USB3 Vision: ~350-400 MB/s, 3-5 m, no grabber - moderate fps, single camera.
• 10GigE: ~10x GigE bandwidth over longer runs - good high-res workhorse.
• CoaXPress CXP-12: multiple GB/s, needs a frame grabber - for high res at high speed.

Lighting uniformity and depth of field

Large sensors and large fields of view make lighting harder, not easier. Illuminating a big inspection area evenly is genuinely difficult: any gradient, hotspot or shadow that a low-resolution camera might tolerate becomes visible across thousands of pixels and can swamp the subtle low-contrast defects you bought the resolution to find. Budget for uniform, well-controlled illumination - large diffuse, bar, dome or coaxial sources sized to the field - and consider flat-field correction to compensate for residual non-uniformity and lens vignetting.

Depth of field also tightens. High resolution often pushes you to smaller apertures to keep the whole field sharp, which reduces light and forces brighter lighting or longer exposures - and longer exposures conflict with high frame rates and any part motion. Telecentric optics, controlled working distance and good part presentation all help. Lighting specialists such as DCM Sistemes can provide large-area uniform illumination matched to a high-resolution field, which is usually money better spent than chasing yet more megapixels.

Single high-res vs multi-camera tiling

When one sensor cannot cover the field at the required pixel density, you have two architectures: a single very-high-resolution camera, or several lower-resolution cameras each viewing a tile of the scene. A single camera keeps optical alignment, calibration and cabling simple and gives one coherent image, but the camera, large-format lens and high-bandwidth interface can be expensive, and it concentrates the entire data rate on one link.

Tiling with multiple cameras can be cheaper per pixel, easier to light tile-by-tile, and lets you parallelise data and processing across hosts - but it adds image stitching, per-camera calibration, mechanical alignment and synchronised triggering. For continuous web or roll-to-roll material, a line-scan camera is often the better answer than either: it builds an effectively unlimited-length image from a moving part at high across-web resolution. Choose based on geometry, motion and total cost of ownership rather than headline megapixels.

Processing cost and total system view

Acquisition is only half the system. A 25-50 MP image at speed can overwhelm a CPU-only pipeline, so high-resolution inspection frequently needs GPU acceleration, region-of-interest processing to avoid analysing every pixel, and careful memory and storage planning if images are logged for traceability or audit. Account for this compute and storage cost up front; it often dominates the running cost of a high-resolution cell.

High-resolution data is also where deep-learning inspection earns its place - subtle, variable or hard-to-rule defects across a large field are exactly the cases where a trained model such as Rosepetal outperforms fixed thresholds, and ROI-based inference keeps it tractable at these data rates. Where the answer spans camera, optics, lighting, interface, frame grabber and software as one integrated system, treating it as a full integration problem (the remit of an integrator such as AIS Vision) avoids the classic failure mode of a perfect sensor starved by the wrong lens, link or lighting.