Computer-vision defect detection brings consistent, fast, fatigue-free inspection to the production line — surface defects, assembly and completeness, weld characteristics. This guide covers what it reliably detects, the sensor and lighting setup that makes or breaks it, the dataset and annotation strategy for the defect-scarcity problem every plant faces, and edge deployment at line speed. It also draws a firm honesty boundary: a vision model surfaces candidate indications mapped to a recognised vocabulary — it does not assign a certified grade.
Last reviewed: June 2026
AI visual quality inspection is the use of computer-vision models to detect defects and verify quality on manufactured parts directly on the production line. Cameras capture each part under controlled lighting, and a model — often a deep-learning detector or an anomaly detector that learns what “good” looks like — flags surface defects, missing or misassembled components, and other deviations at production speed. It delivers consistent, fatigue-free screening that augments human inspection. Critically, it surfaces candidate indications for the quality process; it does not, by itself, constitute certified inspection, metrology, or grading.
Human visual inspection is variable, fatigues, and does not scale to full-rate, 100% inspection. Computer vision offers the opposite profile: consistent, tireless, and able to inspect every part at line speed. That is the case for AI visual quality inspection — not to remove human judgement, but to apply a consistent first pass to every unit and free skilled inspectors for the cases that genuinely need them.
But computer-vision inspection succeeds or fails long before the model. The defect must be made visible to the sensor (lighting and optics), the part must be presented consistently (fixturing), and the system must learn from a dataset that reflects real defects (annotation and the defect-scarcity problem). Teams that treat this as a pure modelling exercise — buy a camera, train a network — consistently underperform teams that invest first in lighting, presentation, and data.
This guide follows that order of importance: what vision can reliably detect, how to capture the image, how to build the dataset, how to deploy at the edge, and — because it is the line that separates an honest capability from an overclaim — exactly what the model’s output does and does not constitute, with welding as the worked example.
Computer-vision inspection spans a spectrum of difficulty and risk. The pragmatic path starts with the unambiguous, binary checks and expands toward the harder, judgement-laden ones as the programme matures. The four categories below, with an honest note on where each sits.
Scratches, dents, pitting, corrosion, contamination, discolouration, coating and paint defects, porosity on cast or machined surfaces. Surface inspection is the most mature computer-vision application — high-contrast defects on a consistent surface are well-suited to learned detectors, while low-contrast or texture-dependent defects (matte, reflective, or patterned surfaces) demand careful lighting and dataset design.
Presence/absence checks (is the connector seated, the screw fitted, the label applied?), correct orientation, correct component variant, and count verification. These checks are often the highest-ROI starting point: they are unambiguous, the failure is binary, and the cost of a missing component escaping the line is concrete.
Surface-visible weld characteristics — undercut, spatter, surface porosity, irregular bead profile, incomplete fill, visible cracks. A vision model can surface candidate indications and map them to a recognised imperfection vocabulary. It is critical to be precise about what this does and does not constitute — covered in detail in the honesty-boundary section below.
Edge, gap, flush, and feature-position checks. Here vision overlaps with — and is often outperformed by — dedicated metrology (laser, structured-light, CMM). Vision is well-suited to fast in-line geometric screening at production rate; certified dimensional measurement remains the domain of calibrated metrology equipment.
The most common reason a vision-inspection project fails is not the model — it is that the defect was never made visible to the sensor in the first place. Lighting, optics, and presentation are the foundation. Get them right and a modest model succeeds; get them wrong and the best model in the world has nothing to work with.
The order of investment: lighting first, optics second, presentation third, model last. This sequence is the opposite of where most teams instinctively spend — and reversing it is the single highest-leverage decision in a vision project.
Lighting is the single biggest determinant of inspection success — more than the model, more than the camera. The defect must be made visible to the sensor. Different defects need different illumination: diffuse dome lighting for reflective surfaces, low-angle (dark-field) lighting to throw scratches and dents into relief, backlighting for silhouette and presence checks, coaxial lighting for flat specular parts. Get the lighting wrong and no model can recover the signal.
Key Decisions
Tooling
Sensor resolution must resolve the smallest defect of interest with margin — a defect must span several pixels to be reliably detectable. Choices span area-scan vs. line-scan (for continuous web/sheet or cylindrical parts), monochrome vs. colour (colour only when the defect is colour-dependent), lens selection and working distance, and frame rate matched to line speed. The optics are specified backward from the smallest defect and the part travel rate.
Key Decisions
Tooling
Consistent part presentation reduces the variation the model must learn. Stable positioning, repeatable orientation, and controlled part-to-camera geometry mean the model sees the part the same way every time — turning a hard, pose-invariant problem into a tractable one. Where presentation cannot be controlled, the dataset and model must absorb that variability, which raises the data cost.
Key Decisions
Tooling
Supervised learning wants abundant, balanced, well-labelled examples of every class. A production line gives you the opposite: overwhelmingly good parts and very few defects. This inversion shapes the entire data strategy — and is why so much of the craft of inspection AI lives in the data, not the model.
A well-run line produces overwhelmingly good parts, so real defect images are rare — the inverse of what supervised learning wants. This drives several strategies: collecting defects over time, deliberately sampling known-bad parts, synthetic defect generation and augmentation, and anomaly-detection approaches that learn ‘good’ and flag departures (so few or no labelled defects are needed to start).
How defects are labelled determines what the model can do: image-level classification (good/defective), bounding boxes (where), or pixel-level segmentation (exact extent, needed for sizing). Annotation must be consistent — disagreement between annotators on borderline cases is one of the biggest hidden sources of model error. A clear defect catalogue and labelling guide, agreed with quality engineers, is foundational.
Map every labelled class to the plant's quality vocabulary and, where one exists, the relevant standard's imperfection terminology. This makes model output legible to inspectors and auditors and keeps the AI aligned with the language the quality system already uses — rather than inventing a private taxonomy no one downstream understands.
Because defects are rare, raw accuracy is a misleading metric — a model that calls everything ‘good’ can score 99% and catch nothing. Evaluation must centre on recall on defects (escapes are the costly error), precision (false rejects waste good parts and operator trust), and per-defect-class breakdowns. The accept/reject threshold is a deliberate business decision balancing escape risk against false-reject cost.
Unsure whether your defects are even detectable with vision, or how to start with almost no defect images? Hyperion runs a focused discovery sprint that assesses defect detectability, designs the lighting and capture approach, and produces a pragmatic dataset and deployment plan for your line.
Quality inspection is a real-time, in-line function. The model runs at the edge — beside the camera, at production speed, independent of the cloud. These are the decision points every vision-inspection deployment must address to reach and stay in production.
Quality inspection runs at line speed, so inference happens at the edge — on an industrial PC or vision-capable edge module beside the camera. Latency must fit the cycle time, and the system must operate independently of any cloud connection. Models are typically optimised and compiled (quantisation, ONNX/TensorRT-class runtimes) for the target edge hardware to hit the required throughput.
The inspection station cannot become the line bottleneck. Inference time must be bounded and predictable to match the takt time. This shapes model-size choices: a smaller, faster model that hits the cycle-time budget and catches the defects that matter beats a larger model that cannot keep up with the line.
A flagged part must go somewhere — a reject gate, a rework lane, a manual-review station. Every decision should be logged with the image and the model's output so rejects are auditable and the data feeds the next training round. Traceability is both a quality-system requirement and the engine of continuous improvement.
Production drifts: new material lots, tooling wear, lighting ageing, seasonal change. A model that was accurate at deployment degrades silently unless monitored. Track reject rates, confidence distributions, and human-override outcomes; feed confirmed misses and false rejects back as labels. Visual inspection is a living system, not a one-time install.
This is the most important section of this guide, and the one most often glossed over by vendors. A vision model is a powerful screening tool. It is not a certified inspection. Being precise about the line between the two is what separates an honest deployment from a liability — especially in safety-relevant or regulated production.
What the AI does: a vision model inspects the visible surface of a weld and surfaces candidate indications — undercut, spatter, surface porosity, an irregular bead profile, a visible crack. It can map each candidate to the imperfection vocabulary of a recognised standard such as ISO 5817, so its output speaks the language your quality system already uses.
What the AI does not do: it does not assign a certified weld quality level (such as ISO 5817’s B / C / D). A certified grade is a formal determination that depends on more than a surface photo: it requires metrology, subsurface examination where the standard calls for it (typically qualified non-destructive testing), the applicable acceptance criteria, and your qualified Welding Procedure Specification (WPS) — carried out by qualified inspectors per standards such as ISO 9712.
The correct framing: the AI is a fast, consistent first pass that flags candidate indications for the qualified process — not a replacement for it. It tells a qualified inspector where to look, in their own vocabulary; the certified determination remains theirs.
The same principle generalises beyond welding. For dimensional features, vision provides fast in-line screening while certified measurement remains the domain of calibrated metrology. The honest position is consistent across every defect class: the AI accelerates and standardises the screening; the certified determination stays with the qualified process and the people accountable for it.
Hyperion runs visual-AI demos live on this site. In the defect-demo you upload a single photo of a part, weld, or surface and an AI layer previews what an inspection layer could flag; the plant-audit demo previews how AI reads a wider plant/process image.
Honesty boundary: these are single-photo previews shown live for illustration — not a calibrated inspection, a formal audit, or a certified grade. As covered above, a vision model surfaces candidate indications; a certified determination needs metrology, qualified NDT where the standard requires it, and your WPS. Verify any output against your own process and a qualified inspector before acting.
A factual account of the background behind this work — verified facts, not marketing claims.
Hyperion runs two relevant demos live on this site: a plant-audit demo and a defect-demo. In the defect-demo a visitor uploads a single photo of a part, weld, or surface and an AI layer previews what an inspection layer could flag. Both are demonstrated live, each with an honest caveat that the output is a single-photo preview — illustrative, not a calibrated inspection. They show the shape of the capability; they are not a production inspection system.
Founder Mohammed Cherifi spent 17+ years in automotive and embedded systems engineering, including work at Renault-Nissan-Mitsubishi Alliance, Cisco, and ABB. In-line visual quality inspection sits at the intersection of optics, embedded vision, line integration, and quality systems — the exact territory of that background.
Hyperion has built Auralink — an edge-deployed agentic platform with 400+ microservices and approximately 20 AI agents (architecture described in the arXiv preprint 2603.08736; a preprint, not a peer-reviewed publication). The edge-inference and real-time deployment patterns that programme exercises are the same ones an in-line vision-inspection deployment needs.
Hyperion is an AI and edge-architecture consultancy. The engagement is vision-system design, dataset and annotation strategy, model development, and edge deployment — working alongside your quality engineers, the lighting/optics supplier, and your metrology and NDT specialists. Hyperion does not certify inspection results, does not replace accredited testing, and does not assign weld grades. It builds the AI screening layer that feeds your quality process.
The most mature and lowest-risk applications are assembly and completeness checks (is the component present, correctly oriented, the right variant?) and high-contrast surface defects (scratches, dents, contamination, coating defects). Weld surface characteristics and dimensional features are detectable as screening, but with important boundaries: vision surfaces candidate indications and fast geometric screening, while certified weld grading and certified dimensional measurement require qualified inspection, metrology, and your welding procedure. The right scope starts with the unambiguous, binary checks and expands as the dataset matures.
Because a defect that is not visible to the sensor cannot be detected by any model. Lighting design makes the defect visible: diffuse dome lighting tames reflective surfaces, low-angle dark-field lighting throws scratches and dents into relief, backlighting handles silhouette and presence checks, coaxial lighting suits flat specular parts. Teams that under-invest in lighting and over-invest in model complexity consistently get worse results than teams that do the opposite. Lighting and optics are the foundation; the model sits on top of it.
Yes — defect scarcity is the normal situation, not a blocker. A well-run line produces mostly good parts, so labelled defects are rare. The pragmatic approaches are: anomaly detection that learns what ‘good’ looks like and flags departures (few or no labelled defects needed to start), deliberate sampling of known-bad parts, synthetic defect generation and augmentation, and accumulating real defects over time to graduate to supervised detection. Most programmes start with presence/absence and anomaly screening precisely because they do not require a large defect dataset.
No — and this boundary is non-negotiable. A vision model can surface candidate indications on a weld's surface (undercut, spatter, surface porosity, irregular bead, visible cracks) and map them to the imperfection vocabulary of a standard such as ISO 5817. It does not assign a certified quality level. A certified grade is a formal determination that depends on metrology, subsurface examination where the standard requires it (typically non-destructive testing), the applicable acceptance criteria, and your qualified Welding Procedure Specification (WPS) — performed by qualified inspectors. The AI is a fast, consistent first pass that flags candidates for that qualified process; it is not the qualified process.
Traditional (rule-based) machine vision uses hand-engineered algorithms — thresholds, edge detectors, template matching, blob analysis — and excels at well-defined, high-contrast, deterministic checks (measure this gap, confirm this feature). Learned (deep-learning) vision excels where defects are variable, low-contrast, or hard to specify with explicit rules — surface texture defects, subtle cosmetic flaws, variable appearance. They are complementary: many production systems use rule-based vision for the deterministic metrology-style checks and learned models for the fuzzy, appearance-based defect classes.
On the line. Quality inspection runs at production speed, so inference happens at the edge — on an industrial PC or vision edge module beside the camera — with bounded, predictable latency that fits the cycle time and operates independently of any cloud connection. Models are optimised and compiled for the target edge hardware to meet throughput. Aggregated results and images may be sent centrally for monitoring and re-training, but the accept/reject decision is made locally, in real time.
Not by raw accuracy — because defects are rare, a model that passes everything can score 99% and catch nothing. The metrics that matter are recall on defects (an escape is the expensive error), precision (false rejects waste good parts and erode operator trust), and per-defect-class performance. The accept/reject threshold is a deliberate business decision that trades escape risk against false-reject cost, and it should be set with quality engineers, not buried in the model.
No. Hyperion's scope is the AI screening layer: vision-system design, dataset and annotation strategy, model development, and edge deployment. Certified inspection, accredited testing, non-destructive examination, metrology, and weld grading are performed by your qualified inspectors, metrology team, and accredited bodies. Hyperion works alongside those specialists — the AI flags candidates faster and more consistently; the certified determination remains with the qualified process.
ISO (2023). "ISO 5817: Welding — Fusion-Welded Joints in Steel, Nickel, Titanium and Their Alloys — Quality Levels for Imperfections."
Context: Defines quality levels (B/C/D) and the vocabulary of weld imperfections (undercut, porosity, cracks, incomplete fusion, and others). A vision model can map candidate surface indications to this vocabulary; assigning a quality level is a qualified determination, not a model output.
ISO (2007). "ISO 6520-1: Welding and Allied Processes — Classification of Geometric Imperfections in Metallic Materials."
Context: The reference classification and numbering system for weld imperfections, underpinning the terminology used in ISO 5817. The shared vocabulary for mapping candidate indications.
ISO (2022). "ISO 17636 / ISO 17638: Non-Destructive Testing of Welds (Radiographic / Magnetic Particle)."
Context: Standards for non-destructive examination of welds. Cited to mark the boundary: subsurface weld assessment requires qualified NDT, which a surface-vision model does not perform or replace.
ISO (2021). "ISO 9712: Non-Destructive Testing — Qualification and Certification of NDT Personnel."
Context: Specifies the qualification of personnel who perform and certify non-destructive testing — the qualified human process that a certified determination depends on.
BIPM / JCGM (2012). "International Vocabulary of Metrology (VIM) and Guide to the Expression of Uncertainty in Measurement (GUM)."
Context: The metrology reference for certified dimensional measurement and measurement uncertainty — the basis for the distinction between fast vision screening and certified metrology.
Bergmann, P. et al. (2019). "MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection."
Context: IEEE/CVF CVPR. A widely used industrial-inspection benchmark for unsupervised surface-anomaly detection — the canonical reference for the 'learn good, flag departures' approach to defect scarcity.
Hyperion Consulting (2026). "arXiv preprint 2603.08736: Autonomous Edge-Deployed AI Agents for Physical Infrastructure."
Context: Hyperion founder's preprint (not peer-reviewed) covering edge-deployed agent architecture. The edge-inference and real-time deployment patterns are directly applicable to in-line vision inspection.
Whether you are assessing whether your defects are even detectable or planning a full in-line cell, the early decisions — lighting, optics, presentation, dataset strategy — determine success long before any model is trained. Hyperion brings 17+ years of automotive and embedded-systems experience alongside production work in edge-deployed AI, and an honest line on what AI inspection can and cannot certify. Start with a conversation.
Founder & AI Strategy Lead
Mohammed Cherifi is the founder of Hyperion Consulting, with 17+ years in automotive and embedded systems engineering. He specialises in physical AI deployment — bringing operational experience from Renault-Nissan-Mitsubishi Alliance, Cisco, and ABB to computer-vision inspection, edge inference, and industrial AI architecture.
Sensor-driven failure prediction — the companion shop-floor AI capability
From simulation to production-grade robot autonomy, with the same edge-inference discipline
The 6-layer Physical AI Stack for robotics, edge AI, and industrial automation
ISO 26262 and IEC 62443 considerations for edge AI in industrial environments