Skip to Content
Machine Vision · QC · IoT

HarmonyOS × domestic NPU × 2026 vision foundation models
Eyes for the production line — it sees, thinks, and learns on its own.

We package open-source HarmonyOS, domestic vision NPUs, and the most advanced vision foundation models into an end-to-end quality-inspection pipeline — from 16K line-scan cameras and edge inference boxes, to natural-language-promptable VLM inference, all the way to closed-loop linkage with MES/SCADA/security centers. It spots the defect and explains why.

6 domestic NPUs 20+ frontier models HarmonyOS 5.0 unified device layer 2 field cases
HarmonyOS × NPU × Vision Foundation Model
WHY

Why Build Machine Vision as IoT

Traditional machine vision is a one-off project — retrain when the product changes, drift when the lighting changes. We turn it into a self-evolving Internet of Things.

Pain · Rigid Rules

Each category needs thousands of samples, and every product change means halting the line to retrain. When a new defect appears the model doesn't recognize it, so you fall back on scheduled manual inspection.

Turning Point · Language-Promptable

YOLO-World and Grounding DINO turn the detection target into a single sentence: "find scratches, find oil stains." With VLM on the edge, workers converse with the model.

Result · Self-Learning Fleet

HarmonyOS turns each camera into a node; missed-detection samples flow back to the center · auto fine-tuning every week · versions pushed to the whole fleet.

HARDWARE

Domestic Vision Embedded NPU Matrix

From 5 TOPS entry-level to 128 TOPS automotive-grade · all domestic chips · 3 of them have completed official HarmonyOS Std-5.0 adaptation, the rest supported via our BSP porting.

Domestic Vision NPU Matrix

RK3588

Workhorse
Rockchip
6 TOPS HarmonyOS Std-5.0 ✓
8-core A76+A55
RKNN-Toolkit2
8K decode / 6 MIPI

RK3576

Value
Rockchip
6 TOPS HarmonyOS Community Port ✓
RK3588 cost-reduced
Transformer operators
Low power 3-5 W

AX650N

Strong ViT
AXera
18 TOPS Community Port
INT8/INT4 dual precision
Pulsar2 compiler
Native Transformer

SG2300X

Multi-Stream
SOPHGO
32 TOPS Box-side Linux
32-channel H.264 decode
TPU-MLIR
PCIe / USB3

Journey6

Auto-Grade
Horizon Robotics
128 TOPS QNX / Linux
BPU Nash architecture
AEC-Q100
BEV / DETR native

A311D2

Entry
Amlogic
5 TOPS HarmonyOS Std-5.0 ✓
Dual NPU VIP9000
Acuity toolchain
Low BOM / 4K

Recommended models: lightweight lines pick RK3576/A311D2 · multi-class defects pick RK3588 · heavy models + multi-stream pick AX650N / SG2300X · safety fencing + automotive-grade pick Journey 6.

MODELS · 2026

Frontier Vision Model Matrix

Four lines advance in parallel: real-time detection, segmentation, VLM, and anomaly detection — all deployed quantized on edge NPUs.

2026 Frontier Vision Models

Real-Time Detection

From fixed classes to open vocabulary · RT-DETR true real-time
YOLOv11YOLOv12 · AttentionRT-DETRv3YOLO-World v2PP-YOLOE+

Segmentation

SAM family on the edge · one click to segment
EfficientSAM 10MEdgeSAM 9.5MFastSAMMobileSAM v2SAM 2 Video

VLM · It Talks

Line workers ask in plain language — it sees, it answers
Qwen2.5-VL-3BMiniCPM-V 3.0InternVL2.5Moondream2PaliGemma 2

Anomaly Detection

No defect samples needed · zero-shot go-live
EfficientAD 5msDinomaly Zero-ShotPatchCoreSuperSimpleNetAnomalyGPT
The 2026 Leap

The real inflection point is natural-language-promptable edge quality inspection: a worker types "find scratches and oil stains," and YOLO-World / Grounding-DINO-Edge respond zero-shot; 3B-class VLMs (Qwen2.5-VL / MiniCPM-V) run directly on RK3588 / AX650N, generating a descriptive report for every defect image. Synthetic-defect Diffusion + fleet feedback compress go-live from weeks to days.

ARCHITECTURE

Three-Layer Collaborative Architecture

Model · NPU · HarmonyOS — one toolchain all the way to the field

Three-Layer Architecture
CASE 01

Zhejiang Seehoo · Textile Defect Detection

The textile customer's chronic headache is fabric defects: holes, weft skew, oil stains, color deviation, and spots — a manual inspector covers at most 3,000 m a day, and the miss rate spikes once eyes tire. With Seehoo, we replaced traditional inspection on three production lines: 16K line-scan cameras + RK3588 edge boxes + HarmonyOS 5.0 unified reporting + YOLOv12s + EfficientAD dual-model fusion.

  • 6 defect classes · holes / weft skew / oil stains / color deviation / spots / end-out, full coverage at 40 m/min line speed
  • 7-day cold start · YOLO-World zero-shot labeling + synthetic-defect Diffusion first, then backfill with real samples
  • MES linkage · defect coordinates fed straight back to the host system for auto meter-marking / scrapping / traceability
  • Continuous evolution · missed samples flow back weekly, model versions pushed to the fleet via HarmonyOS OTA
"A single inspector used to cover 3,000 m a day and still missed defects; now one line runs 24/7, the miss rate is pushed below 0.3%, and it paid for itself in 3 months."
—— Seehoo Production Lead
Seehoo Textile Defect Detection
Ski Simulator: Electronic Fence + Vision Fence
CASE 02

Beijing Ski Simulator · Electronic Fence + Vision Fence

An indoor ski simulator runs at high speed; a user falling, sliding off the board, or wandering into the tail end can all cause serious accidents. The traditional approach only uses emergency-stop buttons and light curtains, with no awareness of the "on the machine but losing balance" scenario. For the Beijing customer we built a triple-decision system: an electronic fence (safety light curtain + mmWave radar) + a vision fence (stereo camera + YOLO-World + pose estimation), all centrally controlled by HarmonyOS 5.0.

  • Triple fence · light-curtain millisecond cutoff + mmWave coarse range + vision fine pose
  • YOLO-World + pose · four event types: fall / out-of-bounds / third-party intrusion / abnormal posture
  • Decision within 80 ms · RK3576 edge inference + HarmonyOS soft bus driving the motor brake
  • Graded response · alert = audible-visual warning · danger = slow down · critical = emergency stop
"8 weeks of stress testing · 0 false stops · camera to light curtain to motor, one system handles it end to end. From acceptance to go-live at the second branch took just 3 weeks."
—— Beijing Customer · Safety Manager
WORKFLOW

From the First Image to Mass-Production Go-Live · 6 Steps

In DevEco we package this chain into a CI/CD pipeline — 7-day cold start.

STEP 01

On-Site Capture

On-board camera + HarmonyOS HDF unified capture format · automatic logging

STEP 02

Labeling · Prompting

Traditional labeling ✕ + YOLO-World zero-shot masking + T-Rex Label visual prompting

STEP 03

Training · Synthesis

Few-shot fine-tuning + synthetic-defect Diffusion · 7-day cold start

STEP 04

Quantized Deployment

Pulsar2 / RKNN / TPU-MLIR quantize to INT8/INT4 with one command

STEP 05

Edge Inference

On HarmonyOS devices · NPU inference + VLM anomaly explanation

STEP 06

Continuous Learning

Missed samples flow back · weekly auto fine-tuning · fleet-wide model versioning

SCENES

All These Sites Can Go Live

Same architecture · swap the model · swap the light source · swap the HarmonyOS HDF driver, for fast replication.

Textiles · Dyeing & Printing

Holes / weft skew / oil stains / color deviation · high-precision line scan · MES linkage

Safety · Fencing

Light curtain + vision + mmWave triple decision · fall / out-of-bounds detection

3C · PCB

Solder joints / missing parts / polarity / scratches · multi-light switching · sub-pixel level

Food · Agriculture

Foreign-object rejection · appearance grading · ripeness assessment · packaging verification

Turn the Site into an IoT that Sees, Thinks, and Self-Learns

Give us a production line, an image, or a video clip — we package domestic NPUs, the HarmonyOS edge layer, and 2026 frontier vision models into a deployable solution: 7-day cold start, 3-month closed-loop go-live.

Machine Vision · IoT

HarmonyOS × Domestic NPU × 2026 Vision Foundation Models

Eyes for the production line — it sees, thinks, and learns on its own.

Hero

Domestic NPU Matrix

Hardware

RK3588 / RK3576 · AX650N · SG2300X · Journey 6 · A311D2 —— 5-128 TOPS coverage.

2026 Model Matrix

Models

YOLOv12 · YOLO-World · SAM family · Qwen2.5-VL · EfficientAD / Dinomaly.

Three-Layer Architecture

Arch

Case · Zhejiang Seehoo Textiles

Seehoo

16K line scan + RK3588 + YOLOv12s + EfficientAD · 6 defect classes · miss rate < 0.3% · payback in 3 months.

Case · Beijing Ski Simulator Fencing

Ski

Light curtain + mmWave + vision triple decision · RK3576 edge inference · motor linkage within 80 ms · 0 false stops.

Book a Production-Line POC

FAQ

What can the machine-vision inspection system detect?

Deep-learning-based surface-defect recognition, dimensional measurement, OCR, and assembly verification, with millisecond inference.

Which production-line scenarios does it suit?

Suited for in-line inspection on electronics, auto-parts, and pharmaceutical-packaging lines.

How do you deploy it onto an existing line?

We support integration with existing lines / MES. Contact Qingdao Huo15 (phone 18554898815, email support@huo15.com) for model selection and a solution.