Frequently Asked Questions

Yes. The system identifies the reference of each part before picking it, among those you have shown it, and the robot only grips what matches the current order. References can therefore be mixed in the bin without prior sorting.

Through 3D measurement and the part's appearance: it recognizes the visible face and determines how to grip it, including when the part must be placed in a specific orientation. This top/bottom recognition is one of the points we validate on your actual parts.

It selects a part whose tilt and gripping surface allow a secure pick, and that can be reached without hitting the bin or neighboring parts. A part that is tilted too steeply or buried too deep is left for later: the pile rearranges itself as parts are picked.

Yes. The same 3D measurement used to guide the robot also makes it possible to check that the part is compliant before picking it. A non-compliant part does not move on to the rest of the process.

By reducing the time the robot spends waiting for the vision system. Psycle's robot guidance projects target short vision cycle times: the scene is computed while the robot is working, not between two picks. The overall cycle time also depends on the robot, the gripper and the part.

Yes, that is one of the goals of the solution: handling a wide variety of formats and designs with minimal camera adjustments. A new part is presented to the system without rebuilding the cell from scratch.

Anything that is not metallic but remains visible: fiber, plastic shard, piece of gasket, insect. The camera is placed after filling, before the container is closed, and the affected product is ejected; an alert is raised if the defect recurs. Vision therefore complements the metal detector rather than replacing it, and what is detectable on your products is confirmed during the feasibility study.

Yes. A crease under the lidding film, product caught in the seal or an off-center lid can be seen in the image, before showing up in a leak test. Every tray or pouch is inspected at line speed, rather than just a sample.

Yes. The system spots a mention that is missing, offset, illegible or incorrect, as well as a label that is missing, skewed, peeling or belonging to another recipe. This is a key point for allergens: the leading cause of recalls in the food industry is labeling, not microbiological. A drifting coder is flagged from the very first products, not after an entire pallet.

Yes, that is a starting requirement. The housings, optics and lighting are chosen for jet cleaning (IP69K rating) and disinfectants, and to keep a usable image under condensation as well as frost. The mounting leaves the line accessible to operators.

The processing time is matched to your actual line speed, which can reach several hundred products per minute. Format or recipe changes during the day are planned for from the prototype stage. When a new format arrives, your teams update the model themselves, without going back to a service provider.

The location depends on the defect you are looking for. Placed too early, the station misses what happens afterwards; placed too late, it rejects a product that has already been packaged.

• foreign bodies: after filling, before the container is closed;
• sealing: at the outlet of the sealer or tray sealer;
• date and batch number: right after the coder;
• label and allergens: after the labeler;
• filling and capping: after the filler and capper;
• product appearance: on the conveyor, before packaging.

Ejection is linked to your PLC.

It inspects the four areas where defects are concentrated: the cap or closure (presence, crimping), the label (position, creases, bubbles, tears), the marking (legibility of the code and date) and the body and base (fill level, base deformation, foreign body in the liquid). Cosmetic defects, such as dirt, are inspected with the same cameras. The exact list is set with you, based on your bottles and your acceptance criteria.

Yes. The camera measures the liquid level on every bottle, at line speed, and rejects those that are below or above the set threshold. The inspection is contactless, so it neither slows down nor stops the conveyor; it complements sampling inspection, which only sees part of the production.

With an image of the neck, which first checks that a cap is present and matches the expected product reference, then inspects its crimping. Millimeter-scale defects, such as a skewed or poorly crimped cap, are visible in the image. This inspection prevents a poorly closed bottle from ending up on a pallet.

Yes. The system checks the label position (height and rotation), detects creases, bubbles and tears, and checks that the printed reference is the right one for the current product. This last point is most useful during a changeover: a label from another recipe is rejected before an entire pallet is affected.

Yes. The system reads the QR code, the Data Matrix and the date, then validates that the marking is legible, that is, sharp enough to be read again later. The result is linked to the unit produced: each bottle can thus be traced along with its inspection, and a drifting marking is flagged from the very first bottles.

Through the choice of lighting rather than the camera alone. A transparent or glossy bottle reflects light in ways that hide defects; suitable lighting (backlight, diffuse lighting or low-angle lighting depending on the area) makes them stand out. The optical setup is validated on your actual bottles during the feasibility study, before any hardware commitment.

Yes, that is one of the goals of the solution: handling a wide variety of formats and designs with minimal camera adjustments. The model learns what is normal for each product reference, and a format change is done by switching recipes rather than by reassembling the station. With PAQ, your teams add a new reference themselves.

It is ejected from the line, and an alert is raised if the defect recurs, so you can correct the cause rather than sort out its effects. The inspection keeps pace with the line and covers 100% of bottles, instead of a sample. The number of cameras required depends on the areas to inspect and the speed: it is sized during the feasibility study.

It detects seam defects and dents, as well as labels that are torn off, ripped, creased or swollen with an air bubble. Each one is best seen from a specific angle: the seam and a torn-off label from above, a rip and a dent from the side. The camera setup is therefore chosen according to the defects you want to catch.

It depends on the camera resolution and the area observed. For a defect to be represented by at least 2 pixels, a 5-megapixel camera on a cylindrical can 100 mm in diameter can see a seam defect down to 0.1 mm. The finest seam defects, spikes just a few hundredths of a millimeter in size, require a higher resolution or a narrower field of observation.

A front-facing camera is often enough to see a dent, because it breaks the contrast. On the sides, it is harder to spot, especially on an unlabeled can with a shiny surface. If the dents to detect are small, up to six cameras may be needed, spaced every 60° around the can. The exact number is set during the feasibility study.

By teaching the system what your cans look like. A crease is harder to see than a tear and can be mistaken for a printed pattern: the system must therefore be shown several production orders, so that it can tell the difference between a crease or an air bubble and a design that looks similar. A tear is easier to spot when an area of the can is left unlabeled, because the contrast there is strong.

Yes, but it is more demanding. On a shiny, unlabeled can, reflections make dents hard to see on the sides. The choice of lighting and the number of cameras compensate for this; they are validated on your actual cans before any hardware commitment.

Yes, that is one of the goals of the solution: inspecting a wide variety of formats and designs at high speed with minimal camera adjustments. Each new design is learned from production examples, which is what keeps the system from mistaking a pattern for a defect.

Data sovereignty is a strategic issue that Psycle places at the heart of its solutions. As an ISO 27001-certified French machine vision company, we guarantee that your sensitive production data never leaves the perimeter you control, with no transfer to non-European infrastructure. For manufacturers concerned about their technological independence, this combination of European roots and recognized certification is a condition of trust as much as a competitive advantage.

At Psycle, machine vision is not meant to replace the operator, but to equip their expertise. The fine-tuning of a system always comes from bringing together artificial intelligence and your teams' knowledge of the shop floor. This collaborative approach ensures more relevant inspections and stronger buy-in from operators, who remain the decision-makers on what matters.

In sectors such as nuclear, defense or food, a system's opacity is not acceptable: every inspection decision must be traceable and auditable. That is why Psycle rules out the black-box effect and makes quality formulas and indicators explicit. You have a usable history of all inspections, essential for your audits and continuous improvement initiatives.

A custom solution is designed around your real defects, your products and your line constraints, whereas an off-the-shelf solution imposes generic settings. In practice, Psycle starts from your shop-floor expertise to define specific inspection rules, which avoids the blind spots a standard system cannot handle (decorated packaging, reflections, unusual geometries).

Machine vision means enabling a machine to “see” and decide in place of the human eye, but without turning it into a black box. At Psycle, we apply it to your products and processes to systematize your quality inspections, while gradually letting you take control of your systems. The goal is not only to detect a defect, but to understand its causes so you can act at the source.

Yes. The gantry is installed above an existing conveyor, without modifying the process. The camera, lighting and housing are designed for the dust, humidity and vibrations of the line.

Their shape no longer says anything about the original product, so recognition relies on material, texture and color. The model is trained on images from your own flow, including the borderline cases your operators debate.

The location depends on the question you are asking:

• at the infeed, to know what you are receiving;
• after a separator, to measure what it lets through;
• before baling, to qualify the bale.

Several gantries make it possible to monitor each sorting stage.

A frozen model gradually loses accuracy. That is why the accumulated images are used to retrain it as the flow evolves. With PAQ, your teams collect, annotate, retrain and compare versions themselves, for example when a new material or a new supplier arrives.

It largely complements it. Where sampling takes a few kilos, the gantry analyzes everything that passes, day and night, including during shift changes. It continuously tracks the composition of the flow, contaminants and particle size.

No. Many projects usefully stop at the characterization gantry: you finally know the real purity of your output, the losses in the reject stream and the efficiency of each separator. Controlling ejectors or a robot can come later, based on the images already collected.

No. The system works on what it sees: it measures the position, orientation and dimensions of each object in 3D. New product references, mixed pallets and deliveries that do not follow the announced pattern are handled without creating a product record.

Yes, that is precisely its role. The scene is measured before each pick instead of being inferred from a palletizing pattern. The robot takes into account the actual height of the stack, its tilt and any overhangs. It picks the packages in the order that keeps the stack stable.

The path is part of the decision, just like the pick itself. The calculation takes into account the conveyor, fixed obstacles, the neighboring stack and the package being gripped, so that the movement neither knocks anything over nor hits anything.

A package that has fallen between two conveyors, a pallet that is askew or out of gauge, torn film, a burst carton, an item left behind in a tote, an unreadable code. The model learns from your images what your normal flow looks like, and every alert is stored with its photo.

Yes. Commissioning includes the connection to the robot and the PLC, the upload of information to the WMS and the handling of cases that the cell rejects and sends back for manual handling.

Yes, as long as the information is still present in the image. The system reads codes that are damaged, partially hidden or misoriented, then matches the package it has read with the order. Tracking therefore no longer stops at the last successful scan.

Yes. Every step runs on the installed hardware: image capture, processing, decision and archiving. No external connection is required. Updates go through the channel defined by the facility operator, and the model is retrained on site, offline.

On your site, and nowhere else. Production images, annotations and models stay with you. They are not used to train a model shared between customers and do not pass through any third-party service. No telemetry is sent.

Each result is stored with the image that produced it, the version of the model used and the date. A scrapping decision, a sorting decision or an alert can therefore be reviewed and explained months later, to the facility operator as well as to a client.

The model mainly learns what a normal part looks like, then flags anything that deviates from it. It therefore needs far fewer images than an approach that would require many examples of defects. Nor do you have to damage parts to build a training dataset.

Everything is prepared upstream. The mounting, optics and model are validated on a test bench. The hardware is declared in advance, and the intervention is planned to fit within the scheduled access window. Components are chosen knowing that they will not leave the area again, and the settings are designed to hold from one intervention to the next.

Yes. The feasibility study starts from the images you already have, or from representative non-active parts. It establishes what is detectable before any hardware is committed to the installation.

Yes. In fact, it is one of the founding principles of PAP. Psycle's Python SDK for machine vision lets your developers design their own applications directly in Python, building on every component of the software: hardware management, processing flows, business rules, communication with plant systems (MES, ERP, PLCs). The code is readable in plain text, with no black box: your teams control every decision the application makes, evolve it at their own pace and retrain their models without depending on an external service provider.

PAP and PAQ are two complementary software products, designed to work together. PAP supervises and controls production in real time, on the line: it analyzes images, applies quality rules and triggers operational actions. PAQ takes a step back to manage, analyze and improve your entire fleet of vision systems over time: data visualization, model retraining, management of the versions deployed machine by machine. One acts, the other drives improvement.

Yes. PAP offers an operating mode without an HMI (“headless” mode) in which it communicates directly with the line's PLCs and equipment via standard industrial protocols (PROFINET, EtherNet/IP, OPC UA). This mode is particularly well suited to automated cells where no operator station is planned, or to architectures where the PLC orchestrates the triggering of acquisitions and the reception of results. The HMI remains available for supervision and configuration, but is not required for production operation.

That is precisely what PAP's Configuration module makes possible without outside intervention. Switching the active recipe, changing the parameters of a quality rule or adjusting detection thresholds is done directly from the interface, by your teams, without recompiling a single line of code. For more structural changes (a new AI model, a new business rule), SDK mode lets your developers work directly on the application. Either way, your vision stations evolve at the pace of your production, not at the pace of a service provider's availability.

PAP integrates into existing automation architectures in line with industrial standards. It communicates with the line's PLCs and equipment via proven field protocols: PROFINET, EtherNet/IP and OPC UA. To feed data up to your information systems, PAP interfaces with your business tools (an MES such as VIF or an ERP such as SAP) so that inspection results feed directly into your KPIs, with no re-keying and no proprietary connector.

No. PAP is designed to be run by a production operator, not a vision expert. Changing a recipe, adjusting a parameter or triggering an action from the HMI requires no image processing or programming skills. Vision expertise comes in upstream, when the application is developed, whether by Psycle's teams in project mode or by your own developers via the SDK. Once the application is deployed, your shop-floor teams stay fully in control.

A smart camera processes an image and returns a result: that is its strength, and also its limit. As soon as a business rule falls outside the scope planned by the manufacturer (a new product reference, an additional quality criterion, an action to trigger on the line), the room for maneuver runs out. PAP is complete machine vision software: it controls the cameras, applies quality rules your teams can configure and continuously triggers operational actions. The inspection logic remains readable and adjustable, without going back to the vendor for every change.

The results and statistics produced by PAQ are meant to feed your quality and production KPIs, not to stay isolated in a separate platform. PAQ interfaces with standard business tools (an MES such as VIF, an ERP such as SAP) so that the monitoring data from your vision stations flows directly into your existing dashboards, with no re-keying and no proprietary connector. Integration protocols are defined according to your IT architecture during deployment.

PAQ is technically standalone. In the vast majority of deployments, it works with PAP, supervising its applications and improving its models. But nothing prevents you from connecting PAQ to existing vision stations outside this ecosystem, depending on the case. If you are starting from an installed base, the right question is not “PAP or no PAP” but “what data can PAQ retrieve from your current stations”. A quick assessment is enough to answer it.

PAQ imposes no fixed limit on the number of stations monitored. The platform is designed to support deployments at the scale of an industrial group, with multiple sites and dozens of machines. The Administration module allows fine-grained management of access rights according to your organization's structure (each team, site or entity accesses its own scope with no visibility into the others). Before any deployment, a study phase maps your installed base, defines the rights structure suited to your organization and plans a gradual, controlled rollout.

Production images are sensitive data, and protecting them is an integral part of PAQ's design. The platform is available as a Cloud version hosted in France, as well as an On-Premises version for industries where confidentiality is critical (nuclear, defense, aerospace), which can then keep their data within their own infrastructure. This requirement is part of a certified security approach: Psycle is ISO 27001 certified, with full traceability of data and access.

PAQ is technically standalone and can supervise vision stations developed outside the Psycle ecosystem. In practice, it is most often deployed with PAP, with which it integrates natively. If you have an existing installed base running other software, contact us directly: we are in regular discussions with other machine vision software vendors, and we can quickly tell you whether an integration with your current solution is feasible.

Depending on a service provider for every change is not inevitable: it often stems from the lack of a tool that makes these operations accessible in-house. PAQ gives process engineering and quality managers visibility over their stations (deployed versions, performance, drift) and the tools to act: annotating new images, retraining and deploying updated models. A new product, a tightened quality criterion or a line modification become routine operations, managed by your own teams, not projects to outsource.

General-purpose MLOps platforms are designed to manage model training cycles in a data science context. They bring structure, but ignore the realities of a fleet of machine vision systems in production: the link with the stations on the line, managing the application versions deployed machine by machine, or the ability to find one specific image among millions to understand a drift. At Psycle, PAQ is built around these constraints: it is a monitoring platform designed for vision systems that run continuously, not for data pipelines in a lab.

Yes, provided the pouch carries a readable identifier (QR code, Data Matrix, best-before date). The system reads the code, links the inspection result to the identifier and sends the information to the MES for complete unit-level traceability.

Fine creases on transparent film, partial delamination and micro-leaks with no visible trace are the most demanding cases. They require low-angle or structured lighting and AI models trained on real defects, not just simulations.

By archiving unit-level inspection images and results, the system makes it possible to correlate a drift in defects with a change in a process parameter: sealing temperature, forming pressure, film batch. This is one of the key functions of PAQ, Psycle's quality monitoring platform, available in the Cloud hosted in France or On-Premises.

A gradual drift in forming is visible in the inspection data well before it reaches the reject threshold. The system raises an alert as soon as the distribution of measurements leaves the nominal range, so the machine settings can be adjusted in time.

This is one of the most common use cases. By configuring a recipe dedicated to the new format, the system collects inspection data from the very first batches and enables fast statistical validation, without waiting for feedback from the field.

Yes. Vision systems operate in real time without creating a bottleneck on the line. The acquisition frequency is matched to the machine rate during the optical study phase; at Psycle, this step comes before any hardware procurement.

Vision inspects the continuity, width and position of the seal on 100% of units. It raises an alert as soon as an incomplete, missing or out-of-tolerance seal is detected, without waiting for a sampling inspection.

The benefit for your integrator lies in the durability of the integration. The vision system communicates with the PLC via standard industrial protocols (PROFINET, OPC UA, ETHERNET/IP) and sends pick coordinates to the robot controller when needed, without imposing a proprietary architecture. Above all, the vision modules (even highly specific ones) are encapsulated in a standard way in the software: the processing complexity stays on Psycle's side and does not end up in the PLC. Updates can therefore be deployed without involving the integrator or reworking the robot program, a decisive point for long-term maintenance.

The model you choose directly affects the budget. A turnkey system, including hardware and integration, is a larger investment than integrating the software alone into an existing installation. Supplying your own hardware or bringing in your integrator can therefore adjust the cost. Psycle prices both scenarios so you can decide based on your budget and resources.

The effort required from your teams remains limited and focuses on what only you can bring: your product expertise, to define what a defect is. Development, configuration and integration are handled by Psycle. Image annotation can be carried out by your operators, who know your products best, or delegated if you prefer a turnkey project, without the quality of the result depending on it.

The duration of a vision project depends mainly on the complexity of the need: the number of stations to equip, the difficulty of the defects or the task, and the level of integration with the rest of the line. A targeted inspection on a single station is deployed much faster than a multi-station installation integrated into a complex process. Psycle sets a realistic schedule at the end of the assessment, once the scope has been clarified.

Either way, the core of the solution is the software. Running on an industrial PC, it acts as the conductor: it controls the cameras, applies the inspection rules and triggers the decisions. This architecture allows Psycle to supply either a complete system or the software module alone, with no change in the reliability of the result. The surrounding hardware may vary, but the vision logic stays the same.

You don't have to commit blindly to a vision project. The process starts by qualifying the need and removing the technical uncertainties, before defining a scope and a budget. Psycle structures this upstream phase precisely so that you make an informed decision, once feasibility has been established, not before.

A product change does not mean starting from scratch. Psycle's software is designed to adapt quickly: when a new product reference is launched, new inspection rules are defined so it can go into production as soon as possible. You absorb changes to your product range without replacing your vision system.

False rejects decrease over time thanks to annotation. When a compliant product is wrongly rejected, your teams can flag and annotate it: the system incorporates this case and refines its decision for all vision stations. At Psycle, every correction benefits all your lines, which gradually lowers the false positive rate.

A 100% inspection can be trusted provided the system is calibrated on your real cases and its decisions remain verifiable. At Psycle, reliability comes from inspection rules defined on the basis of your expertise and from transparent indicators, never from a black box. You stay in control of what is deemed compliant, which makes automated inspection trustworthy in production.

A machine vision system detects five main families of defects, whatever the industry: cosmetic defects (scratches, stains, deformation, dents, corrosion), conformity defects (presence/absence of a component, assembly or orientation error, incorrect label or marking), dimensional defects (out-of-tolerance dimensions, misalignment, positioning), sealing, crimping or welding defects (leak-tightness, hermeticity), and the presence of foreign bodies. These inspections apply equally to food packaging, a metal part, an electronic assembly or a component in a critical environment — all areas covered by Psycle.

A vision-based quality control system combines controlled lighting, one or more cameras and decision software, installed directly on the line. Each time a product passes, the image is compared with the conformity criteria agreed with your teams, and the verdict comes in a few milliseconds, fast enough to trigger an ejection without slowing down the line.

The difficulty is almost never seeing the defect: it is distinguishing it from the product's normal variations. That is what is at stake during optical qualification, before a single line of code is written, and what separates a system that is reliable in production from a successful demonstration in the lab.

This speed of adaptation is decisive in contract packaging, where the variety of products and the frequency of changeovers make conventional reprogramming too costly. By adapting to a new product in around thirty minutes, the Psycle solution makes it possible to handle varied production runs without sacrificing productivity. Flexibility no longer comes at the expense of throughput.

The benefit of masked-time processing is direct: it reduces, or even eliminates, the time the robot would spend waiting for a decision from the vision system. On a timed operation, every second saved per cycle is multiplied across the entire production. It is one of the levers through which Psycle's solutions use AI to reduce robot cycle times, without changing the robot itself.

The difference lies in where the camera is placed. With on-board vision, the module is mounted on the robot arm: it only sees what surrounds the gripper, and you have to wait for the movement to finish before observing anything else — so you depend on the cycle time. With remote vision, the camera is separate from the robot: it observes wherever you want, whenever you want, independently of the movement in progress.

3D bin picking covers two main configurations: picking parts from a bin or a pallet — typical of depalletizing — and picking loose products directly from a conveyor, common in contract packaging. Psycle designs these applications to absorb variations in both the product and its presentation without heavy reprogramming: a new product reference is handled within a few dozen minutes, which makes it a realistic solution for lines with frequent format changes.

In industry, 3D vision is mainly used to guide robots on tasks that 2D cannot handle: picking parts from a bin, depalletizing, or locating a product whose position varies with each cycle. Psycle's applications use this real-time localization so that machines adapt to unpositioned products and variable environments, without jigs or mechanical fixturing.

Beyond localization, vision guidance is mainly about gaining performance and flexibility: a robot that can see is able to handle loose products, absorb variations in position and reduce its cycle times. Psycle's guidance solutions rely on AI so that machines adapt to variable environments with a fluidity close to human movement, where a conventional robot would require perfect positioning.

Industrial 3D vision covers the technologies that capture the depth of a scene, in addition to its surface, so that an automated system can measure, inspect or guide a robot precisely. It relies on specialized cameras and three-dimensional reconstruction algorithms applied in production environments.

2D vision analyzes a flat image, in width and height, which is enough to read a code or inspect a surface. 3D vision adds depth, which is essential for locating a part in space, measuring surface relief or guiding a robot in a bin of loose parts.

The four main families are stereo vision, structured light, time of flight (ToF) and laser triangulation. Each offers a different trade-off between accuracy, speed and working distance, to be chosen according to the target industrial application.

It allows the robot to locate a part whose position is not known in advance, for example loose in a bin, and then to calculate a suitable gripping path in real time, without dedicated positioning tooling.

That is one of the major benefits of SDK mode. Because you control the code, you adapt your applications at the pace of your lines: a new product, a tightened quality criterion, a protocol to integrate. Your teams are kept informed of the SDK roadmap and trained on an ongoing basis, so they can adopt new capabilities as they become available.

Yes. Applications developed with the SDK run on PAP and integrate seamlessly with PAQ, the platform for monitoring and continuously improving your fleet of vision systems. Your teams visualize data, centralize AI training and manage all of their machines, regardless of how each application was developed.

SDK mode is designed for organizations that have (or want to build) in-house development skills: manufacturers who want to control their applications end to end, and integrators and machine builders who want to industrialize their own vision deployments. These skills are often more accessible than they seem: automation engineers, used to the logic of production lines, generally pick up a language like Python fairly quickly. If you prefer to entrust the design to experts, project mode meets the same need by a different route — and switching from one to the other remains possible.

You stay in control of your code, which remains readable in plain text: nothing locks you in. Psycle's role shifts toward publishing the SDK, advising and training. Support, not dependency. You decide the level of support you need, from co-developing the first machines to full autonomy for your teams.

No, that is exactly what the SDK takes care of. The complex building blocks (acquisition, AI model execution, PLC communication) are already encapsulated: your developers work in Python on the business logic of your application, not on the low-level plumbing of vision. A series of guided exercises and support from Psycle's teams let them build their skills gradually, with no advanced prerequisites in image processing.

The Psycle SDK relies as much as possible on the major open-source computer vision libraries. But where these libraries process an image (one input, one result as output), the Psycle SDK is the code that runs PAP, a vision system software integrated into production. Developing with the Psycle SDK is not just about analyzing images: it means controlling cameras and connected hardware, communicating with PLCs and plant systems (MES, ERP), running AI models and connecting them to your retraining platforms. You are not coding an isolated image processing routine: you are developing a complete quality control application, ready for production.

It detects product in the seal, seal defects and channel leaks, among others. These defects are numerous and varied: the AI learns from the specific characteristics of each one rather than applying a single threshold.

Yes. A poorly sealed lid is visible in the image as soon as it leaves the tray sealer, before it shows up in a leak test or at the customer's site. Every tray is inspected at line speed, not just a sample.

Yes, under the lidding film as well as on the surface of the product. A foreign body can be recognized by a different shade or shape, which lighting reveals down to four hundredths of a millimeter. Detection remains limited to visible bodies: what is detectable on your products is confirmed during the feasibility study.

Marks, dents and clouding stand out when the lighting sweeps across the tray wall. The optical setup is therefore designed to bring out these surface features rather than simply illuminate the part.

Yes, it is a filling defect. The system must recognize the different planes of the image to distinguish product that overflows from product that is properly inside the tray. It learns this difference from images of your production.

Yes, that is one of the goals of the solution: inspecting a wide variety of formats and designs at high speed with minimal camera adjustments. The model learns what is normal for each product reference, and your teams can add a new one using the PAQ quality platform.