AI for Manufacturing: Using Computer Vision for Quality Inspection

Computer vision for quality inspection works best at the edge: cameras on the line, a small AI computer beside them, and decisions made in milliseconds without sending video to the cloud. The hardware is no longer the hard part. An NVIDIA Jetson board, a compact PC with a professional GPU, or even a mini PC with an NPU can run an inspection model. What decides success is the camera and lighting, a model trained on your own good and bad parts, and a pilot that measures real escapes on your line. The same edge approach suits retail analytics such as people counting and heatmaps, but stores also have to meet PDPA obligations for camera footage.

Which edge hardware fits which vision job?

Source: NVIDIA Jetson Orin, Orin Nano Super and Jetson Thor pages (fetched 9 Nov 2026); NVIDIA RTX PRO desktop GPU specifications; ASUS NUC 16 Pro press release (9 Mar 2026). Prices listed 9 Nov 2026. The performance figures use different units (INT8 TOPS, FP4 TFLOPS), so they can't be ranked on one scale.

Why run vision at the edge instead of the cloud?

Three reasons: speed, bandwidth and control. A reject gate has to act while the part is still in front of it, which leaves no time for a round trip to a cloud region. Several high-resolution cameras running all shift produce more video than most factory links should carry. And many manufacturers would rather not send images of their products and processes off site. Running the model beside the line keeps all three under your control, and the line keeps working if the internet doesn't.

For a store, the same logic applies at smaller scale. An on-site box can count visitors and produce heatmaps, then send only the totals to head office.

What does a factory inspection setup need?

Start with the image, not the GPU. Consistent lighting, a fixed camera position and the right lens decide whether a defect is visible at all. Industrial cameras connect through standard interfaces:

• GigE Vision: Ethernet-based, with "uncompromised data transfer up to 100 meters" over Cat5e/Cat6 and about 125 MB/s at 1 GbE. Version 2.0 added 10 GbE and link aggregation.
• USB3 Vision: up to about 350 MB/s according to Basler, for cameras close to the computer.
• MIPI CSI-2: direct camera connectors on Jetson boards. NVIDIA lists up to six cameras on AGX Orin's 16-lane connector.

If cameras draw power over Ethernet, size the switch's PoE budget before you buy; we explain how in PoE budget planning for cameras and access points.

How accurate is AI defect detection?

Published numbers are impressive but don't transfer directly to your line. The PatchCore method (CVPR 2022) reported image-level AUROC "of up to 99.6%" on the MVTec AD benchmark, a public dataset of industrial defects. NVIDIA reported in May 2023 that Pegatron reached "99.8% accuracy on its AOI systems, starting with small datasets", a vendor-reported figure. Neither tells you how many defective parts will slip past on your conveyor with your lighting.

A benchmark score tells you a model can work. Only a pilot on your own line, with your own good and bad parts, tells you it will.

Plan a pilot that collects real images, includes rare defects, and measures both missed defects and false rejects. Keep a person reviewing borderline calls until the numbers are stable.

How many cameras can one edge box handle?

NVIDIA publishes frames per second, not stream counts, so any camera count is an estimate. In NVIDIA's DeepStream 9.2 performance tables, the PeopleNet detector ran at 147 FPS on AGX Orin and 349 FPS on AGX Thor, processing a 1080p H.265 sample with rendering off. As a CALC, 147 FPS covers about four to five cameras inferring at 30 frames a second, or about 14–15 cameras at 10 frames a second. Retail counting rarely needs 30 frames a second; high-speed inspection may need more, plus higher resolution, which lowers the count.

A compact PC with a professional GPU suits sites that also want to retrain models on new defects without shipping images elsewhere. If you are weighing an NPU against a GPU, our post on why NPU and GPU TOPS figures mislead explains why the headline numbers don't compare. For a mini PC with discrete graphics, see the ASUS NUC 15 Performance with NVIDIA RTX.

Can a store run computer vision without cloud processing?

Yes. People counting, queue length, dwell time and heatmaps can all run on an in-store box or an AI NVR, sending only aggregates to head office. Our post on the ASUS NUC in a smart retail store covers the wider store setup.

Edge processing reduces what leaves the store, but it doesn't remove PDPA obligations. PDPC's advisory guidelines (revised May 2024) state that "An image of an identifiable individual captured in a photograph or video recording is personal data about that individual." Stores should put up notices so people have "sufficient awareness" of cameras, though exact camera positions need not be revealed. Staff captured by the same cameras must be told the purpose. Data counts as anonymised only if there is no "serious possibility" of re-identification, and PDPC's 2022 guide on biometric data covers security uses, not shopper analytics. Our view, which is not a PDPC ruling, is that analytics which discard frames on the device and keep only counts reduce what you store, while the cameras themselves still capture personal data. See office CCTV and PDPA in Singapore and Singapore's AI governance requirements. This article is general information, not legal, medical or compliance advice.

What should a pilot look like?

• One station, one defect type: pick a visible, costly defect and a single camera position.
• Fix the image first: lighting, lens and mounting before any model work.
• Collect your own data: good parts and as many real defects as you can, labelled by your QC team.
• Measure escapes and false rejects: run alongside human inspection before switching anything off.
• Plan for change: new products, suppliers or lighting will need retraining.

Training or retraining vision models happens on a workstation, not the edge device; the sizing logic is similar to the one in our architecture and engineering workstation guide.

Buying in Singapore

We list the Jetson Orin Nano Super developer kit at S$696, the Jetson AGX Orin 64GB developer kit at S$5,871.60 and the Jetson AGX Thor developer kit at S$9,075.60 (listed 9 Nov 2026). For training and multi-stream inference, the RTX PRO 4000 Blackwell SFF is S$5,640, and for stores the Synology DVA1622 AI NVR is S$1,390.80 and the ASUS NUC 16 Pro is S$4,281.65. For camera networks, the FortiSwitch 124F-FPOE (24 PoE+ ports, 370 W) is S$1,938. TP-Link VIGI cameras such as the VIGI C350 are price on request. We don't list industrial machine-vision cameras, lighting or rugged industrial PCs; ask us and we'll source to your line's requirements. Prices change; ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Do I need a cloud service for AI quality inspection? No. Inspection models can run on an edge device beside the line, which keeps latency low and images on site.

Which Jetson should a factory start with? The Orin Nano Super suits a one- or two-camera pilot. AGX Orin or AGX Thor suit production stations with several cameras or larger models.

How accurate is AI defect detection? Benchmarks report figures above 99%, but those come from public datasets or vendor reports. Only a pilot on your own parts shows your real escape rate.

How many cameras can one Jetson handle? NVIDIA publishes frames per second, not camera counts. In our illustrative calculation, AGX Orin running PeopleNet handles about four to five cameras at 30 frames a second.

Does in-store people counting fall under PDPA? Video of identifiable people is personal data under PDPC's guidelines, so notices and protection still apply even if only counts are kept.