Optics and lighting
How the lens and the light decide what the camera can see. Most inspection problems are solved, or lost, here.
Field of view (FOV)
Field of view (FOV) is the area of the scene that a camera and lens capture in one image, usually stated as its width and height at the product.
The field of view is chosen together with the smallest detail to be found: the camera's pixels are spread across it, so a larger view means each pixel covers more of the part. It must also leave a margin for variation in product position, or parts drift out of the image. A large product with small defects may need several cameras, a line-scan camera or several images rather than one wide view.
See: Inspection technology Dimensional measurement
Related terms: Spatial resolution, Working distance
Working distance (WD)
Working distance (WD) is the distance from the front of the lens to the surface being inspected.
Together with the lens focal length, it sets the field of view, and it decides how much space the station needs above or beside the conveyor. A short distance can leave too little room for lighting, guards or product changeovers; a long one needs a longer lens and more light. Dome and low-angle lights often have to sit close to the part, so working distance and lighting are planned together at the site survey.
See: Inspection technology Machine vision lighting guide
Related terms: Field of view, Depth of field
Depth of field (DOF)
Depth of field (DOF) is the range of distances from the lens over which the image stays acceptably sharp.
It matters when products vary in height, sit at different positions on the conveyor or have features at several levels. Closing the lens aperture increases depth of field but lets in less light, which calls for a brighter light or a longer exposure, and a longer exposure blurs moving products. Parts that move outside the depth of field give soft edges, which can look like defects or hide them.
See: Reducing false rejects Inspection technology
Related terms: Exposure time and motion blur, Working distance
Spatial resolution (smallest detectable feature, pixel resolution)
Spatial resolution is the size of the area on the part that one camera pixel covers, and it sets the smallest feature a station can detect or measure.
It follows from the field of view divided by the number of pixels across the sensor. A defect or character detail must cover several pixels to be found reliably, not just one, so the camera is chosen from the smallest detail that matters rather than from its pixel count alone. Focus, optics, lighting contrast and motion blur also limit what is resolved in practice, which is why the result is confirmed on your own samples.
See: Dimensional measurement Web & line-scan inspection
Related terms: Field of view, Exposure time and motion blur
Telecentric lens
A telecentric lens is a lens that keeps the magnification of an object constant when its distance from the lens changes, within the lens's working range.
It matters for measurement: with a standard lens, nearer features look larger, so a part that sits higher or tilts reads as a different size. Presence, label and code checks rarely need one. The limits are practical: the field of view is no larger than the front of the lens, so the lens must be at least as large as the part; it is longer and heavier and needs more mounting space; and it works best with lighting matched to it, typically a backlight for silhouette measurement.
See: Dimensional measurement Inspection technology
Related terms: Backlight, Edge and caliper measurement
Backlight (backlighting, transmitted light)
Backlighting is a lighting method in which the light sits behind the part, facing the camera, so that opaque parts appear as dark silhouettes.
It gives sharp outlines for measuring edges, holes and gaps, for finding flash and short shots, and for checking presence and position. On clear containers the light passes through, showing fill level, particles and bubbles. It shows nothing of the top surface. It needs access behind the part, which not every conveyor allows, and dust on the light shows up as a false defect. Labels block the view of the fill level, and foam or condensation can confuse it.
Bright-field lighting (ring light, bar light)
Bright-field lighting is a lighting geometry in which light reflected from a flat surface enters the lens, so the surface looks bright and features that scatter light look darker.
Ring lights around the lens and bar lights above the part are the usual forms. It is the simplest and brightest option and suits matte parts: printed cartons, paper labels, presence and colour checks. On glossy parts it creates hot spots, reflections of the light itself, which can hide print or cause false rejects; moving the light off the camera axis often moves them out of view.
See: Machine vision lighting guide Label & print inspection
Related terms: Dark-field lighting, Dome lighting
Dark-field lighting (low-angle lighting)
Dark-field lighting is a lighting geometry in which light arrives at a low angle so that its reflection from a flat surface misses the lens, leaving the surface dark while scratches, edges and raised features glow.
It is the usual first choice for scratches, dents, burrs, embossed or engraved text and dot-peen codes. Part height and tilt must be consistent, because the effect depends on the angle. Dust and fibres glow as well, and a single bar light shows features running across its light better than those running along it. On clear glass and plastic it shows cracks and chips, which bend light towards the camera.
Dome lighting (diffuse lighting)
Dome lighting is a diffuse lighting method that surrounds the part with even light from nearly every direction, so reflections from shiny, curved or crinkled surfaces merge into an even background.
It is used to read print and codes on cans, foil, blister packs, bottles and other shiny curved packs, and for laser marks on curved metal. The dome must sit close to the part, the camera hole can show as a dark spot on flat mirror-like parts, and because diffuse light hides relief it also hides shallow scratches and dents. Where both print and surface relief matter, two lighting set-ups or two images may be needed.
Coaxial lighting (on-axis lighting)
Coaxial lighting is a lighting method in which a half-mirror, called a beam splitter, sends light down along the camera's own axis, so that flat surfaces square to the camera reflect it straight back into the lens.
A flat, polished face looks bright, while scratches, pits and stains scatter the light and look dark. It suits polished metal, glass, laminated labels, flat foil and laser marks on flat, polished faces. Curved or tilted areas turn dark, so it only works on faces square to the camera, and the beam splitter wastes much of the light, which calls for a brighter source or a longer exposure.
See: Machine vision lighting guide DPM code reading
Related terms: Dome lighting
Line light
A line light is a light that produces a narrow, bright and even line of illumination along the scan line of a line-scan camera.
A line-scan camera exposes each line only briefly, so it needs concentrated light, and the light must be even along its whole length or the image shows bright and dark bands along the material. Light and scan line must stay aligned; a knock to either mounting shows as darker images. Depending on the defect, it is set up as a backlight through the web, as reflected light, or at a low angle.
See: Web & line-scan inspection Machine vision lighting guide
Related terms: Line-scan camera, Encoder
Strobe lighting (strobing, strobed light)
Strobe lighting is a lighting method in which the light switches on only for the camera exposure, triggered together with the camera.
With a short exposure it freezes moving products and can outshine ambient light, because during that moment the station light is far brighter than the room. Light and exposure timing must match: if the flash starts late or ends early, images vary in brightness from product to product. Pulse limits come from the datasheets of the light and its controller and differ between lights.
Polarising filter (polariser, cross-polarisation)
A polarising filter is an optical filter that passes light polarised in one direction, so that a polariser on the light and a crossed polariser on the lens can suppress glare while keeping surface detail.
It works because light reflected from many surfaces is partly polarised. It helps with glare on film, glossy print and wet or coated surfaces. Filters cost light, so the light must be brighter or the exposure longer. Changing the lighting geometry, for example to a dome or an off-axis light, often works better, so polarisers are tried on the samples rather than fitted by default.
See: Machine vision lighting guide Label & print inspection
Related terms: Dome lighting, Bright-field lighting
Shroud and ambient light (enclosure, light shield)
A shroud is an enclosure or shield around the inspection area that keeps ambient light, such as daylight, overhead lights and light from neighbouring stations, out of the camera's view.
Ambient light changes through the day, and a recipe set up in the morning can start rejecting good products when the afternoon sun comes in. A shroud, strobing the station light and a lens filter that passes only the station light are the usual remedies, often combined. A shroud also shields operators from flashing light. It needs to be designed with cleaning and changeover access, or it tends to be left open.
See: Machine vision lighting guide Reducing false rejects
Related terms: Strobe lighting, Drift
Cameras and imaging
The camera types, where the processing runs, and the settings that decide whether a moving product gives a sharp, repeatable image.
Area-scan camera (matrix camera)
An area-scan camera is an industrial camera that captures a whole two-dimensional frame at once, like a photograph.
It is the usual choice for separate products such as bottles, cartons, labels and machined parts: one trigger gives one image per product or view. The frame must cover the product with a margin for position variation, and the product must stay sharp during the exposure. For long or continuous material, or round products imaged while they rotate, a line-scan camera is often a better fit. Several area-scan cameras may be needed to see every face that matters.
See: Inspection technology Web & line-scan inspection
Related terms: Line-scan camera, Field of view
Line-scan camera
A line-scan camera is an industrial camera that images one thin line at a time and builds a continuous image by stacking the lines as the material or product moves past.
It suits webs of film, foil, paper, board, nonwovens, coil and sheet, and rotating cylinders. Each line is usually triggered by an encoder so that the image keeps its scale when speed changes; without one the image stretches or squashes and defect positions along the roll are lost. It needs a bright line light, a stable material path without flutter, and a line rate that keeps up at top speed. Wide webs use several cameras side by side.
See: Web & line-scan inspection Machine vision lighting guide
Related terms: Encoder, Line light, Area-scan camera
Smart camera
A smart camera is an industrial camera with the processor, inspection software and lighting control built into the same housing.
It suits one or two checks from a single view and keeps a station compact, with less hardware in the cabinet. The limit is processing headroom: many tools on one image, deep learning or several views of one product usually point to a PC or vision controller instead. HIKROBOT states that a build of VisionMaster runs on some of its smart cameras; whether a recipe moves between a smart camera and a PC station depends on the model, software version, modules and licence.
Vision controller and PC-based vision (PC-based vision, vision PC)
A vision controller is a purpose-built industrial computer with camera ports and digital I/O that runs the inspection software for one or more separate industrial cameras, and PC-based vision is the same arrangement on an industrial PC.
This architecture suits several cameras, heavier processing such as deep learning, an operator screen and image storage. The trade-offs are more hardware, hardening of the PC for dust, heat and vibration in the plant, and its lifecycle: operating system updates and replacement hardware have to be planned. Camera ports and I/O should leave room for later extensions. The inspection software can be the same as on a smart camera; only where it runs changes.
See: Inspection technology HIKROBOT VisionMaster integration
Related terms: Smart camera, Digital I/O
Exposure time and motion blur (shutter time, motion blur)
Exposure time is the length of time the camera sensor collects light for one image, and motion blur is the smearing that appears when the product moves during that time.
On a moving line a short exposure keeps edges sharp; how short depends on line speed and the smallest detail to be seen, and it is set during the sample test. A shorter exposure collects less light, so it is paired with a brighter or strobed light. Blur softens edges and characters, which can look like a defect, hide one or shift a measurement. An exposure that was right at one speed can be too long after the line is speeded up.
See: PLC & reject integration Reducing false rejects
Related terms: Strobe lighting, Depth of field
GigE Vision and GenICam (GigE Vision, GenICam)
GigE Vision is a camera interface standard for sending images and camera control over Gigabit Ethernet, and GenICam is a standard hosted by the EMVA that gives cameras a generic programming interface whatever interface they use.
Together they let vision software acquire images from compliant cameras of different makes. HIKROBOT states that its GigE area-scan and line-scan cameras are compatible with GigE Vision and GenICam, and that VisionMaster is compatible with the GigE Vision standard. Compliance makes image acquisition possible; it does not mean every camera feature can be controlled from every software, so a mixed combination is confirmed before quoting. Cabling, switches and network load for several cameras are part of the design.
3D vision (laser profiling, laser triangulation, structured light)
3D vision is imaging that measures the height or shape of a surface, for example by laser profiling, where a camera views a laser line projected across the part, or by structured light, where a camera views a projected pattern.
It is needed for flatness, warp, thickness, step height, weld bead shape and dents that do not show in a flat image; a standard 2D camera looking straight down sees one plane. 3D sensors have their own limits: shiny or dark surfaces can return a weak or noisy profile, parts of the shape can be hidden from the sensor, and laser profiling needs the part to move past at a known speed or with an encoder. It is a different station design and costs more.
See: Dimensional measurement Surface defect detection Cost & budget guide
Related terms: Encoder
Inspection methods and software
How an image becomes a pass, a fail or a value: the two families of methods and the tools a recipe is built from.
Machine vision (vision inspection, camera inspection)
Machine vision is the use of industrial cameras, optics, lighting and software to capture images of products on a production line and turn them into decisions or data, such as pass or fail, a measurement or a code read.
In a working station the camera is only one link: a trigger, lighting designed around the defect, a link to the PLC and a reject mechanism matter as much. Vision sees surfaces, outlines, print and marks; it cannot see inside opaque material or behind other parts, and a visible seal check is not a leak test. In electronics the same kind of camera-based check is often called automated optical inspection (AOI).
See: Inspection technology All applications
Related terms: Rule-based inspection, Deep-learning inspection
Rule-based inspection (traditional machine vision)
Rule-based inspection is a machine vision method in which an engineer builds and tunes a chain of tools, such as locate, measure, count and read, and compares each result with a set limit.
Every decision can be traced to a measured value and the limit it was compared with, which makes it predictable and easy to explain in an audit. It fits faults that can be defined as a measurement or a rule: presence, position, dimensions, counts, codes and print in a consistent font. It struggles when good product varies naturally in appearance, because fixed rules then need many exceptions. It depends on stable imaging; when the image changes, the limits no longer fit.
See: Rule-based vs deep learning Inspection technology
Related terms: Deep-learning inspection, Pattern matching
Deep-learning inspection (AI inspection)
Deep-learning inspection is a machine vision method that uses a model trained on labelled images to classify an image, locate objects or mark defect regions, instead of rules written by an engineer.
It suits defects that vary in shape and appearance on naturally variable surfaces: scratches, marks, moulding faults and difficult print. It moves effort from writing rules to collecting and consistently labelling images, validation, and retraining when products or materials change. The model returns a score or a region; people still set the threshold that turns it into pass or fail. It cannot find a defect the image does not show, so lighting and optics come first. HIKROBOT supplies VisionMaster's deep-learning modules as a separate package.
Anomaly detection
Anomaly detection is a deep-learning method trained on images of good products only, which flags image regions that differ from what it learned as normal.
It helps when defects are rare or too varied to collect many examples of each type. It reduces the need for defect images during training, but not during validation: defect samples are still needed to prove the model catches what matters. Because it flags anything unusual, normal variation that was missing from the training images, such as a new print shade or supplier, is flagged too. HIKROBOT's VisionMaster page lists anomaly detection trained on normal samples only, with abnormal areas shown as a heat map (vendor-stated); the V4.3.0 deep-learning package lists it, while the V4.4.0 package listing does not, so availability is confirmed for the version used on each project.
See: Rule-based vs deep learning
Related terms: Deep-learning inspection, Reference sample set
Region of interest (ROI, inspection window)
A region of interest (ROI) is the part of the image that an inspection tool examines, set so that the tool ignores everything outside it.
A tight region keeps backgrounds, guides and neighbouring products out of the decision and shortens processing; narrowing the inspected region is one way to cut false rejects. When products move, the region must follow the part, which is why a locate step usually comes first. A region set too tight lets a shifted label or code fall outside it; one set too wide picks up background variation. In costing, ROI also means return on investment, a different term.
See: Reducing false rejects Label & print inspection
Related terms: Pattern matching
Pattern matching (part location, template matching)
Pattern matching is a machine vision tool that finds a taught reference pattern in the image and reports its position and angle, so that other tools can be placed relative to the part.
It is usually the first step in a recipe. Products do not arrive in exactly the same place, so regions of interest and measurements are positioned from the part, not from fixed image coordinates. The reference must be taught from a representative part: a pattern taught on one unusual part, or on a feature that varies between batches, gives failed locates, which show up as rejects of good product.
See: Rule-based vs deep learning Presence & counting
Related terms: Region of interest, Inspection recipe
Blob analysis (connected-component analysis)
Blob analysis is a machine vision tool that groups neighbouring pixels of similar brightness or colour into regions, called blobs, and measures their number, area, position and shape.
It is used to count objects, check presence and absence, and find spots, holes or contamination that stand out clearly from the background. It depends on a clean threshold between object and background, so uneven lighting and shadows cause miscounts. Where objects touch or overlap in the image, a count can be wrong even with good lighting, so product presentation and spacing matter as much as the tool.
See: Presence & counting Rule-based vs deep learning
Related terms: Region of interest, Backlight
Edge and caliper measurement (edge detection, caliper tool)
Edge and caliper measurement is a machine vision tool that finds edges, where brightness changes sharply, along a line or band in the image and measures the distance between them.
It is the basis of most dimensional checks: lengths, widths, diameters, gaps and positions. The result is only as good as the edge the software finds, so a backlit silhouette gives the most repeatable edges; burrs, chamfers and shadows are agreed in advance, because they change where the edge sits. Converting pixels to units needs calibration with a target, and tilt or movement of the part looks like a change in dimension.
Related terms: Telecentric lens, Backlight, Repeatability and gauge R&R
Inspection recipe (job file)
An inspection recipe is the stored set of settings for one product, covering camera and lighting settings, references, tools, limits and expected codes, that a station loads to inspect that product.
Each product or SKU usually has its own recipe. At changeover the PLC or HMI selects it, the station reports back which recipe it is running, and the PLC checks they match before production restarts; a wrong recipe is a common cause of rejects after a changeover. Recipes should be named, versioned, backed up and changed under an agreed approval, with operator access to limits restricted. Recipe files and backups are part of the handover.
See: PLC & reject integration How we deliver a project Support & changes
Related terms: Pattern matching, Drift
Reference sample set (golden samples, master samples)
A reference sample set is a small, controlled set of known good and known defective products used to check that a station still gives the same results and images as when it was accepted.
It is run at agreed moments, such as shift start, after cleaning and after changeover, and result and image are compared with the stored reference. It catches drift early: ageing lights, a film on the optics, a camera that moved. The samples must be stored so that they do not wear, fade or get mixed into production. A golden sample should be representative rather than the best part on the bench, because references taught on a perfect part cause good products to be rejected.
See: Reducing false rejects Dimensional measurement
Related terms: Drift, Borderline sample library
Codes and marking
Printed and marked codes: reading text, reading symbols, and why reading a code is not the same as grading it.
OCR (optical character recognition)
OCR (optical character recognition) is a machine vision function that reads printed or marked characters and returns them as text.
On a line it is used when the text must be logged, passed to a traceability system or compared with a value held elsewhere. It needs the characters in a predictable area with consistent contrast; inkjet dot characters, white lines from worn thermal print heads and low-contrast laser marks make reading harder. For date and batch codes whose expected text is known, OCV is usually the better check. HIKROBOT describes VisionMaster's OCR as deep-learning based (vendor-stated).
See: OCR/OCV & code inspection Label & print inspection
Related terms: OCV
OCV (optical character verification)
OCV (optical character verification) is a machine vision check that starts from the text that should be printed and confirms that it is present, complete and legible.
For best-before dates and batch codes the line already knows what the code should say, so OCV answers the useful question: is the expected text there and readable? It needs the expected string for every run, ideally sent automatically from the recipe, coder or PLC rather than typed in at each changeover. Its limit is the source: if the recipe and the coder both hold the wrong date, they agree and the error passes.
See: OCR/OCV & code inspection
Related terms: OCR, Inspection recipe
1D barcode (linear barcode)
A 1D barcode is a code that stores data in the widths of parallel bars and spaces, read across one direction.
Retail product codes and logistics labels are the common examples on a line. A reader or camera station decodes the code and checks that its content matches the product or the order. Reading fails when bars are smeared or cut off, when the code sits on a curved or crinkled surface at an unfavourable angle, or when glare covers it. A code that decodes has not thereby been graded; print quality against a formal standard needs a verifier.
2D code and Data Matrix (2D code, Data Matrix, matrix code)
A 2D code is a code that stores data in a pattern of cells in two directions, and Data Matrix is a square or rectangular 2D code widely used to mark products and parts.
2D codes hold more data in less space than a 1D barcode and include error correction, so a partly damaged code can often still be read. Data Matrix is common for part serials and direct part marking; QR codes appear more often on packaging. Reading depends on enough pixels per cell, even contrast and a mark that faces the reader. The content is usually checked against the expected product or serial, and duplicates can be flagged by the PLC or MES.
Direct part marking (DPM)
Direct part marking (DPM) is the marking of a code or text directly onto the surface of a part, by dot-peen, laser or ink-jet, instead of on a label.
It gives a part a permanent identity for traceability. Contrast comes from the mark itself: dot-peen from the shape of the dots, so it usually reads best under low-angle light; laser marks from a change in colour or texture, often read under dome or coaxial light. Marks drift as a stylus wears or laser settings change, and later processes such as heat treatment, blasting or coating can reduce contrast or fill the mark. A read straight after marking gives early warning.
See: DPM code reading Metal components
Related terms: Dark-field lighting, 2D code and Data Matrix
Code reading and code verification (barcode grading, barcode verification, code grading)
Code reading is decoding a barcode or 2D code to obtain its data, whereas code verification, also called grading, measures the print quality of the code against a formal standard with a dedicated, calibrated verifier.
A reader shows that a code decodes and holds the right content; it does not give a formal quality grade, and a code that reads on one reader can still fail a customer's grading. For direct part marks the quality guideline is ISO/IEC 29158. Production camera stations are set up for reading and matching content; if a customer or retailer requires graded codes, that needs a verifier and its own procedure. addaScan stations read codes; formal grading is not offered.
See: OCR/OCV & code inspection DPM code reading
Related terms: 1D barcode, 2D code and Data Matrix
SSCC (Serial Shipping Container Code)
The SSCC (Serial Shipping Container Code) is the GS1 identification key for a logistic unit such as a case or pallet.
It is usually carried as a barcode on the logistic label. On a line, a reader decodes the SSCC at a defined point and the station or the warehouse management system compares it with the unit expected there, so a pallet is released only on a match. Reading confirms that the label decodes and matches; the GS1 Logistic Label Guideline gives placement and verification recommendations, and formal grading of the label is a separate check.
See: Carton & pallet verification
Related terms: Code reading and code verification, 1D barcode
PLC and reject integration
What happens after the decision: signals to the line, tracking the product and making sure the failed one really leaves.
Trigger, latency and jitter (photo-sensor trigger, hardware trigger, trigger jitter)
A trigger is the signal that tells the camera or vision controller that a product is in position and an image should be taken, usually from a photo-sensor that the product breaks as it arrives.
The sensor must see every product once and only once: small gaps between products, transparent containers and wobbling products cause missed or double triggers, which give images of half a product or no result at all. Timing also varies. Latency is the delay from trigger to result; jitter is how much that delay, or the product's position at the moment of exposure, varies from one product to the next. Reject tracking must allow for the full variation, not only the average, and counting triggers against results shows when one goes missing.
See: PLC & reject integration Reducing false rejects
Related terms: Encoder, Shift-register tracking
Encoder (rotary encoder, shaft encoder)
An encoder is a sensor on a conveyor or roller shaft that sends pulses as it turns, so that the line's movement is measured by distance rather than by time.
It is used when line speed varies, when the line stops and restarts, or when products are close together. The PLC can then track a failed product to the reject by distance, and a line-scan camera uses the pulses to trigger each image line so the image keeps its scale. The encoder measures the belt or roller, not the product: where products slip, accumulate or are pushed together, a sensor at the reject point re-identifies the product. The encoder itself must not slip on its shaft or wheel.
See: PLC & reject integration Web & line-scan inspection
Related terms: Shift-register tracking, Line-scan camera
Shift-register tracking (reject tracking)
Shift-register tracking is a PLC method that stores the position and result of each inspected product in a queue and fires the reject when that product reaches the reject point.
With encoder counts, tracking follows distance, so it stays correct through speed changes and stops, provided each product moves with the belt. A simple time delay does the same job on constant-speed lines where products do not slip, but goes wrong when the line slows, stops or restarts. Two checks keep tracking honest: count triggers against results, and keep products from changing order between camera and reject. If the saved image of a rejected product shows a pass, tracking removed the wrong product.
Related terms: Encoder, Reject confirmation, Trigger, latency and jitter
Digital I/O (discrete I/O, hard-wired I/O)
Digital I/O is the set of on/off input and output signals, wired directly between a vision station and the PLC, used for the trigger, pass/fail, result valid, ready and alarm.
It is the simplest route: fast, easy to test with a meter and independent of network settings. A 'result valid' pulse tells the PLC when to read the pass/fail output. The output type, NPN or PNP, must match the PLC input card. Digital I/O carries only a few on/off states, so codes read, measurements and recipe numbers need an industrial network. Where a network protocol does not match the PLC, hard-wired I/O is the usual fallback.
See: PLC & reject integration Inspection technology
Related terms: Industrial Ethernet
Industrial Ethernet (PROFINET, EtherNet/IP, Modbus TCP)
Industrial Ethernet is the family of Ethernet-based protocols, such as PROFINET, EtherNet/IP and Modbus TCP, that PLCs and devices use to exchange data on a production line.
Over a network a vision station can send more than pass/fail: the codes read, measurements, the active recipe, counters and a sequence number per product. Protocol support depends on the camera, code reader or controller model, not on the software brand. HIKROBOT's catalogue lists TCP/IP, Modbus, serial, UDP and EtherNet/IP for VisionMaster, and PROFINET on specific smart code reader models. The exact model is checked against your PLC before quoting; a gateway or hard-wired I/O covers a mismatch.
Reject mechanism (rejector, air jet, pusher, diverter)
A reject mechanism is the actuator that removes a failed product from the line, such as an air jet, a pusher or a diverter that switches products into another lane.
The choice depends on product weight and stability, spacing between products, available compressed air and space on the conveyor. The mechanism must remove the failed product without knocking over its neighbours or taking a good product with it. An air-pressure switch shows when the supply is low and the reject may not work. The reject is only finished when a sensor confirms that the product left the line.
See: PLC & reject integration Inspection technology
Related terms: Reject confirmation, Shift-register tracking
Reject confirmation (reject verification)
Reject confirmation is a check, made by a sensor at the chute, in the bin or just after the reject point, that a failed product actually left the line within an agreed window.
Firing the reject is not proof that the product left. If the product is still on the line, or a good product was removed, the PLC raises an alarm or stops the line, as agreed in the design; a bin-full sensor completes the picture. Without confirmation, a reject that sometimes misses looks like a working station, and a defect that was detected still reaches the customer. Confirmation is tested at SAT on the real line.
Fault behaviour and fail-safe design (fault handling, fail-safe design)
Fault behaviour is the agreed, documented response of a line when part of the inspection chain fails, such as no result in time, a camera offline, low air pressure, a failed reject confirmation or a full reject bin.
Each case needs a decision with production and quality: reject the product, stop the line or raise an alarm. A fail-safe design chooses them so that a fault does not silently let uninspected product through, for example with a heartbeat that lets the PLC notice a silent station, and a logged, access-controlled inspection bypass. This is general control practice, tested at FAT and SAT; it is not a certified machine-safety function, and hazards to people are handled by the machine's safety system.
See: PLC & reject integration How we deliver a project
Related terms: Reject confirmation, Factory acceptance test
Metrics and acceptance
How a station is proved and kept working: the two error rates, measurement studies, samples and the two acceptance tests.
False reject and false reject rate (false call, overkill, false positive)
A false reject is a good product that an inspection station rejects, and the false reject rate is the number of good products rejected divided by the number of good products inspected.
It is measured by running a known number of good products at production speed and counting the rejects, with the sample set, speed and duration recorded. It is reported together with the missed defect rate, because either can be lowered by raising the other. False rejects cost more than the product in the bin: operators stop trusting the station, and sooner or later someone loosens a limit or switches it off. Fix the imaging before touching the limit.
See: Reducing false rejects How we deliver a project
Related terms: Missed defect and missed defect rate, Drift
Missed defect and missed defect rate (escape, false accept, false negative)
A missed defect, also called an escape, is a defective product that an inspection station passes, and the missed defect rate is the number of defective products passed divided by the number of defective products fed.
It is measured by feeding known defective products at known positions and counting how many pass. A small run with no misses shows that detection works on those samples; it does not prove the station misses nothing in long-term production. For critical faults the target and the test method are agreed in writing before FAT. Loosening a limit to cut false rejects raises this rate, which is why limits are changed together with the quality team.
Throughput and cycle time (line speed)
Throughput is the number of products a line or station handles in a given time, and cycle time is the time available for, or taken by, one product or machine cycle.
For a vision station the cycle time is the budget for trigger, exposure, processing, result handover and reject. Heavier processing, such as deep learning or many tools per image, needs faster hardware or more time per product. Line speed and product spacing are gathered at the site survey; where parts must stop to be measured, cycle time decides whether that is possible. Throughput can be written into the acceptance criteria and checked at FAT, as in the envelope line case.
Repeatability and gauge R&R (gauge R&R, gauge repeatability and reproducibility)
Repeatability is the variation in a station's results when the same part is inspected or measured several times under the same conditions, and a gauge R&R study adds reproducibility, the variation from reloading, operators or set-ups.
For measuring stations it is the basis of acceptance: the same parts are measured repeatedly, with reloading, and the spread is compared with your tolerance, which must be wide compared with the station's own variation. For pass/fail checks, a part that passes once and fails the next time shows poor repeatability, often a borderline part or unstable imaging. Agreement with your CMM or gauges is a separate check that finds systematic differences.
Drift
Drift is a slow change in a station's images or results over time, caused by changes in the station or the product rather than by real defects.
Common causes are light output falling with age, a film of dust or oil mist on the lens or cover glass, a camera or guide moved during cleaning or changeover, and a new label stock, resin, ink or supplier. The usual symptom is false rejects creeping back after months of stable running. Reference samples run at agreed intervals, and a chart of good products rejected per shift or batch, show drift early. Fix the cause and re-verify rather than retuning limits on the line.
Sample test (feasibility test, sample testing)
A sample test, or feasibility test, is the imaging of a customer's own good and defective products under candidate cameras, lenses and lighting to find out whether the defects that matter can be seen and separated, before a station is designed or quoted.
It asks "can this be seen and separated?", not "does the station perform on the line?". The samples must represent real production: the normal variation of good product, every defect type and borderline cases, each labelled. The result is a written feasibility conclusion: what was visible, under which conditions, and what still has to be verified on the line. There is no fixed sample quantity; it depends on how much the product and the defects vary.
Borderline sample library (borderline samples, limit samples)
A borderline sample library is a collection of parts that are just acceptable and just rejectable, each recorded with the quality team's decision, a photo and the reason.
It settles arguments on the line, trains operators, and keeps limits and deep-learning labels consistent: if two inspectors disagree about a borderline part, a model learns the disagreement. It is used to set limits during the sample test and to check limit changes later. The library is kept apart from production stock and extended when new materials or suppliers bring new kinds of borderline part.
Factory acceptance test (FAT)
A factory acceptance test (FAT) is the test of a built inspection station against agreed samples and criteria before it is delivered to site.
At FAT the station is checked for detection of each agreed defect on the FAT samples, false rejects on good samples, recipe changeovers and its outputs; the customer may witness it. The result is a FAT record, with any open points agreed in writing. FAT shows that the imaging, the decisions and the signals work on the built station; it does not prove that the reject works on the line, which is why a site acceptance test follows.
Site acceptance test (SAT)
A site acceptance test (SAT) is the test of an installed inspection station on the customer's line, with real production and seeded defect samples, against criteria agreed before the build.
SAT repeats the FAT checks at line speed and adds the reject mechanism, reject confirmation, PLC signals, fault behaviour and the operator workflow, with production, quality and maintenance staff taking part. The SAT report is the basis for handover, training and the support plan. Criteria that were not written down before the build are hard to test fairly at SAT, so they are agreed in the proposal.