Machine vision is often lumped in with the broader field of computer vision, but on a production floor it means something narrower and more disciplined. Industrial machine vision uses one or more cameras, combined with deterministic image processing algorithms, to make automated pass/fail decisions in real time: is the part present, is it the right dimension, is the surface free of defects, is the code readable, and where exactly is the part sitting so a robot can pick it up. Unlike research-oriented computer vision or general AI image recognition, a machine vision system on a line is judged by repeatability, cycle time, and traceability — it has to make the same correct call thousands of times a shift, at line speed, and log the result. That constraint shapes every hardware and lighting decision described below.
Cameras: Area-Scan vs Line-Scan
Most machine vision systems use area-scan cameras, which capture a complete two-dimensional frame in a single exposure, the same way a conventional digital camera works. Area-scan is the default choice for discrete parts on an indexing conveyor, tray, or fixture: the part stops (or is tracked), the camera fires, and the whole field of view is captured at once for measurement or defect analysis.
Line-scan cameras instead capture a single line of pixels at a time, and build a complete 2D image line-by-line as the product moves continuously past the sensor. This is the standard approach for continuous web or roll material — sheet metal, textiles, printed film, paper, glass — where there is no natural stopping point and the product can be arbitrarily long. Line-scan systems require tight synchronization with an encoder on the line so each captured line corresponds to a known increment of material travel; get that wrong and the resulting image stretches or compresses.
Sensors, Resolution, and Field of View
CCD (charge-coupled device) sensors were historically preferred for image quality and low noise, but CMOS (complementary metal-oxide-semiconductor) sensors now dominate the market because they are cheaper to produce, consume less power, and support much higher frame rates and on-chip processing. Modern CMOS global-shutter sensors have closed most of the historical quality gap, so CCD is now mostly a legacy or niche choice.
Resolution and field of view (FOV) are locked in a tradeoff that every vision engineer has to negotiate. A higher-resolution sensor spread across a larger FOV gives more pixels per millimeter of the part, which improves measurement precision and the ability to resolve small defects — but it also means more data per frame, which increases processing time and can force a slower frame rate. Conversely, shrinking the FOV to fit a smaller resolution budget improves speed but limits how much of the part (or how many parts) fit in one image. A common design error is picking a resolution that "looks fine" on a monitor without calculating the actual pixels-per-millimeter needed to reliably detect the smallest required feature — as a rule of thumb, a feature should span at least 2-3 pixels to be measured reliably, and defect detection typically wants more margin than that.
Lenses set the working distance and FOV for a given sensor: focal length determines how far the camera must sit from the part to fill the frame with the desired area, and changing the lens is usually cheaper than changing the sensor when a FOV requirement shifts. Telecentric lenses, which hold magnification constant across a range of working distances, are frequently specified for precision dimensional measurement because they eliminate the perspective error that occurs when a part is slightly closer to or farther from a standard lens.
Lighting: The Most Overlooked Variable
Ask any experienced vision integrator what causes the most failed installations, and the answer is rarely the camera — it is lighting. Ambient light changes with the time of day, nearby equipment, and even operators walking past a cell, and inconsistent illumination forces the software to compensate for variation that shouldn't exist in the first place. A well-designed, enclosed, controlled lighting setup does more to make an inspection reliable than any amount of clever algorithm tuning. The goal of lighting design is to maximize the contrast between the feature you care about and everything else, so the image-processing step becomes almost trivial.
Backlighting
Backlighting places a diffuse light source behind the part, opposite the camera, so the part appears as a sharp black silhouette against a bright, uniform background. It is the standard technique for measuring outer dimensions, hole diameters, and edge profiles, because it eliminates surface texture and color entirely and leaves only a clean, high-contrast outline for edge-detection algorithms to trace.
Bright-Field, Direct, and Ring Lighting
Direct or ring lighting illuminates the part surface from roughly the same axis as the camera, providing even general-purpose illumination for surface inspection, presence/absence checks, color verification, and reading printed text or codes. Ring lights (LED rings mounted around the lens) are the most common off-the-shelf lighting tool in machine vision precisely because they work adequately for a broad range of everyday inspection tasks.
Dark-Field Lighting
Dark-field lighting strikes the part surface at a low, grazing angle rather than head-on. On a flat, defect-free surface, most of that light reflects away from the camera and the surface appears dark. Where a scratch, dent, embossed mark, or particle disrupts the surface, it catches the grazing light and throws a bright highlight or shadow directly at the lens. This makes dark-field the go-to technique for detecting fine surface defects that are nearly invisible under direct light.
Diffuse and Dome Lighting
Diffuse dome lighting surrounds the part with light arriving from every angle simultaneously, eliminating hot spots and shadows. It is the preferred technique for shiny, curved, or highly reflective parts — metal caps, chrome trim, glossy plastic — where a directional light source would otherwise create glare or specular reflections that blind the sensor in exactly the area being inspected.
| Lighting Technique | How It Works | Best For |
|---|---|---|
| Backlighting | Diffuse source behind the part creates a silhouette against a bright background | Outer dimension and edge measurement, hole diameter checks |
| Bright-field / Ring | Direct illumination on the same axis as the camera | General surface inspection, presence checks, code/text reading |
| Dark-field | Grazing-angle light highlights surface irregularities via reflection or shadow | Scratches, dents, embossed marks, fine surface defects |
| Diffuse / Dome | Omnidirectional, shadow-free illumination from all angles | Reflective, curved, or glossy parts prone to glare |
Common Industrial Vision Tasks
Dimensional measurement relies on edge-detection algorithms to locate part boundaries in the image, then converts pixel distances to real-world units using a calibration performed against a known reference target (a calibration grid or gauge block), producing a pixel-to-millimeter scale factor for the specific lens and working distance in use.
Presence/absence detection checks whether an expected feature — a cap, a label, a fastener — exists in the correct location, typically using simple blob analysis or pattern matching rather than precise measurement.
Defect detection looks for surface flaws, contamination, cracks, or cosmetic blemishes, and is often the hardest traditional vision task because "defect" can have high natural variability in shape, size, and appearance.
OCR and OCV (optical character recognition and optical character verification) read and validate printed lot codes, expiration dates, and serial numbers; OCV specifically confirms that a code is not just readable but matches an expected value and meets print-quality grading standards.
Barcode and 2D code reading covers both traditional 1D barcodes (linear stripe patterns encoding limited data, read along one axis) and 2D codes such as Data Matrix, which encode substantially more data in a small square footprint and include built-in error correction, allowing them to be read even when partially damaged or printed directly onto a part via dot-peen or laser etching.
Robot guidance uses vision to locate a part's position and orientation in 2D or 3D space so a robot arm can pick it up correctly, even when the part is not in a fixed, repeatable location. The hardest version of this problem is bin picking, where parts are randomly oriented and overlapping in a bin, requiring 3D vision and pose-estimation algorithms to select a graspable part and orientation for each pick.
Traditional Vision vs. Deep Learning
Traditional rule-based machine vision — edge detection, blob analysis, geometric template matching, pixel counting — remains the default choice for the majority of industrial inspection tasks because it is fast, deterministic, and easy to validate: engineers can point to exactly why a part passed or failed, which matters for quality audits and regulatory traceability. Deep-learning-based vision tools are increasingly used for the harder tail of problems, particularly cosmetic defect classification where the range of acceptable and unacceptable variation is difficult to describe with explicit geometric rules. The tradeoff is real: deep learning models require large labeled training datasets, retraining when the product or lighting changes, and are inherently harder to fully explain or certify than a rule-based algorithm. In practice, most industrial systems default to traditional vision for well-defined pass/fail geometry and reach for deep learning only when rule-based approaches genuinely cannot handle the variability of the inspection — not as a first choice, but as an escalation.