Zebra's NPU Edge Sensors Put AI Anomaly Detection on the Packaging Line

Zebra Technologies has deployed AI anomaly detection for full-garment inspection in sports merchandise, with fixed-mount reading delivering a 30 percent increase in pack rate over handheld scanners, according to Charlie Long, vice president and general manager of machine vision at Zebra. Long laid out the approach in an interview on the Visions: A Machine Vision and Automation Solutions Podcast with host Jim Tatum, Senior Editor at VSD.

The context is seasonal. Championship runs in college football and Major League Baseball create sudden demand spikes for jerseys, hats, collectibles and apparel. Long cited the Los Angeles Dodgers' unexpected path to the World Series as an example of rush orders that suppliers had not anticipated. Each surge raises two risks: counterfeit goods entering the supply chain, and licensed suppliers failing to hold quality at production speed.

What do rule-based systems miss?

Traditional rule-based machine vision struggles with the unknown, Long said. Engineers program edge tools to check label position, dimension patch size, and verify expected misprints or color problems — checks against specifications defined in advance. When demand shifts to an unexpected team, or a new counterfeit enters the market, those fixed rules fail.

AI-based tools invert the workflow. The system examines an entire garment through segmentation rather than inspecting only regions of interest where problems occurred historically, and compares each item against trained known-good examples. Zebra applies this at pack stations and in returns classification, where a customer may return a non-authentic item against an expensive authentic purchase. Long described the capability as 100 percent inspection instead of spot checks on known problem areas.

Counterfeits most often show lower construction quality, differently sized labels or patches, and incorrect placement. Trademarks must appear in the correct place and orientation. The AI system can also compute a composite score: instead of one disqualifying defect, segmentation combines several small anomalies into an overall verdict that flags a suspect garment.

Which hardware does the work?

Zebra attacks the problem on two tiers. For high volume and low deployment cost, the company uses the NS42 Smart Sensor, an NPU-powered AI-on-the-edge sensor with point-and-click setup for anomaly detection. Where accuracy requirements rise, Zebra moves to PC-based systems on the same platform and tool base, including the CV70 shown at Automate, which Long said can put a couple hundred megapixels across a single garment and resolve detail down to a misplaced stitch.

Deployment speed has changed most. Five years ago, Long said, deployment meant days or weeks deciding on specifications, then days or weeks programming them into the inspection system. Now the system trains on known-good samples in a couple of hours. Zebra attributes the gain to new foundational models that understand more of what the vision system sees, rather than requiring exhaustive user training.

The same economics apply on the labor side. Training 25 QA inspectors on a rush production run takes days or weeks to reach proficiency, with higher fallout. The machine vision system reaches production readiness in hours, without added human inspection.

Where is the ROI?

The fastest return comes from pack-bench applications, Long said: the right item in the right box going to the right customer. A packer often cannot distinguish a Clemson Tiger from a USC Spartan. A classification algorithm running on the edge verifies the product at 100 percent inspection accuracy as it enters the box, cutting returns, unhappy customers and bad reviews.

Automation also removes the handheld scanner. Fixed-mount reading, combined with multimodal inspection through machine vision, barcode reading or RFID, makes operators hands-free. Across 100 packers in a facility, Long said, a 30 percent pack-rate increase roughly means every three people yield another effective packer.

How does the system handle new products and judgment calls?

When a championship patch or an unexpected World Series shirt appears, ease of deployment and tool adjustment lets operators adapt quickly. The system also surfaces anomalies no one configured it to find and requests human feedback — the human-in-the-loop model — to adjust its intelligence.

Defect scoring is not a binary go/no-go, Long said. He pushed back on the black-box reputation of AI inspection: Zebra's software shows a clear threshold between good and bad inspections, flags excessive overlap that suggests poor lighting or a mismatched AI scenario, and lets users annotate good and bad images to define which variations matter to their brand. Zebra deploys the same systems at licensed suppliers to verify they meet licensee requirements across the supply chain.

Beyond visible inspection, Long pointed to hyperspectral imaging for contamination and material variation invisible to RGB cameras, plus moisture detection in packaging. In 3D, dimensioning optimizes tote and container fill. Parcel carriers, he said, spend more on shipping empty air than they make on shipping product, an area where 3D measurement makes a measurable difference.

What comes next?

Long said the half-life of advancement continues to decrease, and Zebra has launches planned for the beginning of 2027 tied directly to AI development, covering ease of use and capability in machine vision. Zebra placed on a top-ten list of AI-ready companies, Long said, and has doubled down on that investment.

The open question for manufacturers and distributors is deployment: whether the gap between a couple hundred megapixels of AI segmentation and a traditional rule-based edge tool has already made manual inspection at pack benches — and the handheld scanner — obsolete in high-volume sports apparel.