A regional industrial manufacturer was losing customers to quality escapes. Their inspection team was experienced — but line speed had outpaced what human eyes could reliably catch at volume. We designed a vision system that runs at line speed without slowing it down.
The line runs at the same speed. The defects just don't make it through anymore.
[INDUSTRIAL MANUFACTURER] A regional precision components manufacturer supplying the automotive and HVAC sectors. The plant operates two production lines running 16 hours per day, producing roughly 14,000 components per shift. Their quality specification requires dimensional accuracy within tight tolerances, with surface defects — micro-cracks, inclusion marks, and finish irregularities — classified as field-critical failures in the automotive supply chain.
The company had an established quality inspection team. Their manual inspection pass rate was strong on obvious defects. The problem was subtle ones — defects that were real, were field-critical, but didn't present as obviously wrong to an inspector working at line speed with a visual pass/fail judgment.
Manual visual inspection has a fundamental biological ceiling: the human eye processes at a certain speed, and attention degrades over a shift. At the line speed this plant operated, inspectors were making a pass/fail judgment on a component every 2.1 seconds. For gross defects — a crack that's immediately visible, a surface void that's obvious under normal lighting — the team's detection rate was high. For subtle defects — hairline subsurface cracks, shallow inclusion marks at specific angles, marginal finish irregularities — the detection rate was inconsistent and hard to quantify because escaped defects only appeared as customer returns weeks later.
The customer returns data told the story. One automotive OEM had issued two formal corrective action requests in eighteen months, citing field failures that traced back to escaped quality defects. A third would trigger a supplier qualification review. The financial exposure from a de-listing was significant — this customer represented roughly 31% of revenue.
The plant had evaluated off-the-shelf machine vision systems. None of them could detect their specific defect profile without an unacceptable false-positive rate — flagging good parts as defective, which would require a re-inspection step that would slow the line and increase labor cost. They needed a system tuned to their specific components, their specific defect types, and their specific throughput requirement.
We approached this as a hardware-software co-design problem — not a software-only problem. Off-the-shelf vision systems fail in industrial environments because their cameras and lighting aren't specified for the specific optical challenge. The defect profile here — shallow surface marks that only appear under specific angles of illumination — required lighting design as much as model training.
We ran a two-week optical characterization study of the defect types — capturing 3,000+ images of known-defective and known-good components under different lighting angles, wavelengths, and polarization configurations. The study identified that a combination of coaxial dark-field illumination with an 850nm near-IR band revealed the subsurface micro-crack signature consistently, while leaving good components visually clean. We specified and sourced a line-scan camera at 4096px resolution running at 60kHz line rate — fast enough for the line speed with margin to spare.
Over four weeks of production, we collected 22,000 labeled images — defective components pulled from their existing QA rejection bin (with defect type annotated by their quality engineers) and good components randomly sampled from outbound stock. We trained a convolutional defect detection model with a custom loss function tuned to minimize false negatives (missed defects) even at the cost of a slightly higher false positive rate — a deliberate tradeoff for a field-critical application. The model was iteratively refined through three rounds of production testing before go-live.
The system integrates with the existing line control PLC via digital I/O. When the model flags a component, a pneumatic rejection gate fires within 180ms — fast enough at line speed to divert the part without slowing the belt. Rejected parts go to a re-inspection bin rather than direct scrap, allowing quality engineers to verify model decisions and feed corrections back into the training pipeline. The whole system runs on an edge compute module mounted in the line's control cabinet — no cloud dependency, no network latency risk.
Measured against the full defect taxonomy used in their quality specification, including all three field-critical defect types. Validated through 90-day production audit with dual inspection (system plus manual reinspection on a sample basis).
Measured at 6 months post-deployment versus the 6-month pre-deployment baseline. The automotive OEM issued a formal supplier commendation letter. No further corrective action requests.
The inspection system runs entirely within the existing dwell time at the inspection station. Line throughput is unchanged. No additional cycle time was added to the production process.
Good parts flagged for re-inspection. Slightly higher than the plant's initial target (1%), but acceptable given the catch rate achieved. Re-inspection bin adds less than 4 minutes of labor per shift.
The lighting characterization study at the start — two weeks that felt like overhead — turned out to be the most important part of the project. Every machine vision project we've done that skipped the lighting design phase eventually ran into an accuracy ceiling that couldn't be solved in software. Light the defect correctly first. Then train the model.
The decision to route flagged parts to a re-inspection bin rather than direct scrap created a continuous improvement loop that kept improving model performance post-deployment. Six months after go-live, the model had seen 400+ additional labeled examples from production. The catch rate improved by 0.4 percentage points from the initial deployment figure.
The quality engineers were skeptical at the start — they'd been burned by a previous vision system that generated so many false positives it was effectively turned off. We addressed that by setting a conservative confidence threshold at launch and being transparent about the tradeoff: a slightly higher false positive rate in exchange for a genuinely reliable catch rate. Once they saw the first month of results, they asked us to tighten the threshold. We're now on the third iteration of threshold tuning.
Tell us about your production environment, your defect types, and your throughput requirements. We'll tell you whether visual AI can help, what it would take to build, and what a realistic catch rate looks like for your application.
Book a consultation →