The machine vision image sensor is not the most glamorous component in an inspection system, but it is the one that limits everything else. No amount of image processing or deep learning recovers information that was not captured at the sensor. Getting the sensor right is a prerequisite; getting it wrong means rebuilding the optical stack mid-project. This article focuses on the practical engineering decisions: CCD vs CMOS, how to size resolution for an inspection task, frame rate requirements, pixel size and its implications for dynamic range and noise, and how illumination requirements vary by task. For the broader deployment context, see our manufacturing inspection decision guide. What is the difference between CCD and CMOS for machine vision? CCD sensors dominated industrial machine vision for decades because of their low noise, uniform sensitivity, and high dynamic range. CMOS sensors have largely displaced CCDs in new designs, but the tradeoffs are worth understanding before defaulting to either. Property CCD CMOS (Scientific/Industrial Grade) Read noise Very low (~2–5 e⁻) Low to moderate (~3–10 e⁻ for back-illuminated) Dynamic range High (70–80 dB typical) High in modern sensors (70+ dB) Rolling vs global shutter Global shutter standard Global shutter available (costs more) Power consumption Higher Lower On-chip integration Limited Extensive (ROI, binning, HDR modes) Cost Higher Lower at equivalent resolution Availability Declining Wide availability, active development The practical conclusion from our sensor evaluations: for new machine vision projects, CMOS sensors with global shutter are the default choice (observed pattern across our manufacturing inspection engagements; not a benchmark against named devices). CCD remains appropriate only when legacy system compatibility forces it, or when a specific low-noise requirement — fluorescence imaging is the classic case — cannot be met by available CMOS options. Global shutter is non-negotiable for moving targets. Rolling shutter sensors read out row by row, which causes geometric distortion on parts moving during exposure — the classic “jello effect.” At 1 m/s conveyor speed, a 10ms readout time causes 10mm of apparent skew, which destroys dimensional measurement accuracy and degrades defect detection. How do you size sensor resolution for an inspection task? Resolution sizing starts with the minimum detectable feature size. The standard rule is to have at least 2 pixels covering the smallest feature of interest, though 3–4 pixels provides a more reliable detection margin in our experience, particularly when the feature is low-contrast. Worked example. A part 100mm wide, imaged on a 2/3” sensor (8.8mm × 6.6mm) at full width: Magnification: 8.8mm / 100mm = 0.088× If the sensor is 5000 × 4000 pixels, pixel size projected onto the object is 100mm / 5000 = 20µm per pixel Minimum detectable feature at the 2-pixel rule: 40µm If the minimum defect size is 100µm, this resolution is adequate with margin. If the minimum defect size is 50µm, you need higher sensor resolution or a smaller field of view. There is no third option. Resolution selection checklist Minimum defect size defined in millimetres (not qualitative terms like “small”) Field of view defined (largest part dimension + positioning tolerance) Required pixel pitch at object plane calculated (minimum defect / 2–4) Required sensor resolution calculated (FOV / pixel pitch) Lens resolving power checked against required pixel pitch (MTF specification) Sensor format and working distance verified against available lens options Frame rate, exposure, and the illumination it forces Frame rate requirements are set by throughput: parts per minute divided by the number of cameras imaging the same part. A line producing 120 parts per minute requires the camera to acquire at least 2 frames per second, plus overhead. That sounds trivial — but the binding constraint is almost never frame rate. It is exposure time. At 1 m/s conveyor speed with 50µm pixel pitch, allowable motion blur of one pixel sets the exposure ceiling: 50µm / (1000mm/s) = 50µs. At 50µs the sensor collects very little light, which drives the illumination requirement straight back at you. You need enough light to generate adequate signal in 50µs. Strobe lighting at 5–10× the continuous-equivalent intensity is the typical solution, and LED strobes can hit this at durations well below 50µs when driven with a high-current pulse. Pixel size and its implications Pixel size — the physical size of each photosite on the sensor — affects three things at once: Full-well capacity: larger pixels hold more charge before saturating, which raises dynamic range. Noise floor: larger pixels have lower dark current per unit area, which improves low-light performance. Resolution: smaller pixels enable higher spatial resolution at the same sensor format. Typical industrial sensors have pixel pitches of 3.5–7µm. Sensors with 3.5µm pixels pack more resolution into the same format but show lower dynamic range than 7µm pixel sensors at equivalent total resolution. For inspection tasks with high contrast (dark defect on light background), small pixels and high resolution are the priority. For tasks with low contrast — subtle surface texture variation — or wide dynamic range requirements like shiny surfaces with both specular highlights and dark recesses, larger pixel pitches are preferable. This is one of the few sensor decisions where the right answer genuinely depends on the part you are inspecting. Illumination requirements by inspection task Illumination is inseparable from sensor selection. The right sensor under the wrong illumination produces worse results than a mediocre sensor under optimal illumination. Treat the optical stack — sensor, lens, illumination — as one design problem. Inspection task Illumination type Reasoning Surface scratch detection (specular surface) Dark-field or coaxial Scratches disrupt uniform reflection; dark-field makes them bright against dark background Edge / dimensional measurement Backlight Silhouette gives sharp, high-contrast edge regardless of surface finish Colour / coating verification Diffuse dome light Minimises specular highlights; reveals true surface colour Solder joint inspection Multi-angle structured light Height variation requires angled illumination; 3D reconstruction from multiple angles Barcode / text reading Diffuse or ring light Needs consistent reflectance across the label surface Contamination on dark background Bright-field, coaxial Contamination reflects differently from clean surface In our experience, specular surfaces are the most challenging to illuminate consistently. Metal parts, glass, and polished plastics require either dark-field geometry (to make scratches appear bright) or coaxial illumination (to make surface anomalies appear dark), and small changes in part orientation dramatically affect image appearance. Robotics-fed inspection stations with controlled part orientation produce far more consistent results than conveyor-fed stations with variable orientation — an observed pattern across our deployments, not a benchmarked rate. Spectral considerations Most machine vision uses broadband white light or near-infrared illumination. Spectral selection earns its complexity when: Inspecting transparent or translucent materials — near-IR penetrates some plastics; UV reveals fluorescent contamination. Differentiating materials with similar visible appearance but distinct spectral signatures. Working around surface coatings that absorb specific wavelengths. Monochrome sensors paired with a specific illumination wavelength typically outperform colour sensors for single-parameter inspection: higher quantum efficiency at the selected wavelength, and no demosaicing penalty in the image pipeline before the inspection algorithm runs (whether that algorithm is a classical OpenCV pipeline or a learned model in PyTorch or ONNX). Practical sensor specification process Define minimum detectable feature size and field of view → calculate required pixel pitch. Define throughput and conveyor speed → calculate maximum exposure time. Define surface properties → select illumination type. Calculate required illumination intensity for adequate signal in maximum exposure time. Select sensor format based on required resolution, lens availability, and integration requirements. Verify global shutter availability in the selected sensor. Prototype the optical stack — sensor, lens, illumination — on representative parts before committing to a design. Steps 1–4 are often skipped, with teams selecting cameras based on specifications that sound good rather than requirements derived from the inspection task. This is consistently the path to a system that needs to be rebuilt after integration testing reveals it cannot reliably detect the target defect class. The decision belongs at the start, not after the housings are mounted. FAQ
