Plastic converters are moving from sample-based quality checks to 100% inline vision inspection, a topic tracked by PlasticsToday under the headline "Vision Systems Detect Defects in Real Time." The headline groups extrusion, blow molding, and injection molding together, yet the latency budgets that decide whether real-time vision actually works differ by an order of magnitude from one process to the next.

In a blown film line running at 20–30 m/min, the camera must capture and classify a defect before the winder indexes — a 100–200 ms window. In a rotary-table injection molding cell at 30–60 parts/minute, the camera has 1–2 seconds between indices. The PlasticsToday piece does not specify which application it treats, but that gap is what makes "real-time" a project-specific specification rather than a vendor feature.

What does "real-time" mean in a vision system?

Real-time machine vision means deterministic processing within one production cycle, not simply fast processing. The GigE Vision and CoaXPress interface standards define cable and frame-grabber behavior at the byte level. A CoaXPress 2.0 channel carries up to 12.5 Gbit/s, the bandwidth needed for a 4k line-scan camera reading a 2 m web at 40 kHz line rate. GigE Vision at 1 Gbit/s delivers roughly 120 MB/s, sufficient for an area-scan camera running a 2-megapixel sensor at 60 fps in 8-bit monochrome. Processing latency on top of that transfer time — 30–80 ms on a GPU-accelerated classifier such as the Nvidia Jetson Orin — fits inside either budget.

How does lighting change the verdict?

Diffuse front lighting pulls out scratches and sink marks on opaque polymer surfaces. Dark-field illumination picks up contamination and short shots on transparent films. A structured laser line projects a triangulation profile that converts a flat extrusion surface into a height map with sub-millimeter vertical resolution — the standard method for catching die-lip buildup or profile sag on cast sheet. Wrong choices are common: polarized front lighting on a clear PET preform hides haze that the same camera resolves under diffuse backlight.

Is the pipeline still rule-based, or has deep learning taken over?

Both, in nearly every modern installation. Blob analysis, edge filters, and pixel-intensity thresholds handle gross geometric checks at line speed — a missing tie bar on a molded housing, a short shot on a connector, an edge tear on film. Convolutional neural networks flag the cosmetic defects that resist fixed thresholds: streaks, blush, gels, and weld-line discoloration in clear polymers. A retrained network can drift, however, where a rule fails the same way every shift — a non-trivial point for any plant that runs a single line for weeks between changeovers.

What does this trend mean for compliance?

Audit trails matter as much as accuracy once a reject decision is logged. FDA 21 CFR Part 11 in pharmaceutical packaging, EU GMP Annex 11, and ISO 22000 for food-contact lines all require electronic signatures, access controls, and time-stamped image archives on every quality decision. A vision system that drops an image because the storage buffer filled is no longer just a quality problem; it is a regulatory finding waiting to surface in the next customer audit.

The piece in PlasticsToday fits a broader pattern: real-time defect detection in plastics has moved from a CAPEX-optional upgrade to a baseline expectation on lines serving pharmaceutical, medical, and food-grade converters. Plant engineers comparing quotes will want three numbers from each bidder:

• Line rate at the rated false-reject rate
• Time to retrain the network on a new defect class
• Worst-case storage and network downtime behavior under 21 CFR Part 11 audit

Vendors that answer those questions with measured data rather than marketing claims will win the next round of line rebuilds.