Frontend
Operator workspace
- Six-camera overview, fused view, and focused inspection
- Live runtime state and defect-event updates
- History retrieval and image-level review
- Model, sample, parameter, and alarm governance
CASE / 01Computer Vision · Edge Systems · Model Governance
Real-Time Multi-Camera Surface Inspection
One production chain from acquisition to evidence, review, and model governance.
01 / Overview
RoleVision Pipeline & System Engineer
FocusMulti-camera acquisition / Inference / Traceability / Governance
StackPython, PyTorch, OpenCV, Django, Redis, Vue 3, .NET
ContextConfidential industrial deployment; GPL-3.0-only public frontend rebuilt with synthetic data
02 / Product evidence
A public demo rebuilt from the real system boundary.
These screens preserve the operational structure of live inspection, evidence traceability, and model governance, while replacing all client identifiers, production imagery, business records, and metrics with explicitly synthetic content.
03 / The problem
The engineering problem was larger than training a detector. Six line-scan views had to be acquired without blocking the production path, associated with the correct workpiece and section, processed through preprocessing and defect inference, and turned into durable evidence that operators could inspect and review.
Constraints
01Six-view acquisition and edge bandwidth02Preview and formal detection isolation03Workpiece and section lineage04Rare defects and hard negatives05Durable review evidence04 / System approach
Separate the hot path, business truth, and review loop—then make every boundary observable.
01Edge AcquisitionSIX VIEWS
02Preprocess + ROIMICRO-BATCH
03Anomaly + DefectINFERENCE
04Evidence + ReviewDURABLE LOOP
01
Preview frames travel directly from edge capture to the browser, while formal frames are persisted and acknowledged before asynchronous inference. Runtime state, defect events, image evidence, and review changes remain separate but traceable contracts.
05 / System architecture
Four modules, three data planes, one traceable inspection contract.
The architecture separates runtime control, latest-only preview, and durable formal inspection. The public description preserves verified software boundaries and waiting semantics while omitting client identity, site topology, storage addresses, alarm thresholds, credentials, capacity settings, and production parameters.
Backend
Business truth and orchestration
- Own workpiece, section, and runtime identity
- Acknowledge formal-frame ingress and schedule workers
- Persist frame, image, detection, and review lineage
- Expose REST, event WebSocket, and governance APIs
Capture
Six-view edge acquisition
- Drive six line-scan cameras across two logical edge nodes
- Serve latest-only raw, focused, and fused preview streams
- Gate formal capture with frozen runtime context
- Write original images before submitting metadata
Algorithm
Vision inference service
- Run preprocessing and ROI extraction
- Batch anomaly and defect inference
- Read images from shared storage by location
- Return detections, positions, and evidence metadata
05.1 / MODULE COMMUNICATION MAP
One system, three paths with different guarantees.
Control, preview, and formal inspection have different consistency and latency goals. Separating them prevents operator visibility, business truth, and inference throughput from blocking one another.
Low latency / latest-onlyDurable / traceableControl and business context
Runtime control
Production signalsSignal and identity
Establish the current workpiece, section, and phase
Runtime adapter
BackendBusiness truth
Own runtime state and the frozen context
Runtime control
CaptureCapture gate
Use the complete snapshot to permit formal frames
Live preview
CaptureCamera callbacks
Build per-camera, focused, and fused previews
Memory / JPEG
Edge streamKeep the newest frame
Older views may be overwritten under pressure
Preview WebSocket
FrontendOperator workspace
Consume preview without waiting for inference
Formal inspection
CaptureFreeze formal frame
Bind frame, workpiece, section, and camera identity
Durable write
Shared storagePersist original image
Keep large image bytes in the data plane
JSON ingress
BackendValidate and acknowledge
Queue the path and context, then release capture
HTTP batch
AlgorithmBatch vision inference
Read by location and run preprocess, anomaly, and defect
REST + Events
Backend / FrontendEvidence loop
Persist, emit deltas, and review on demand
Key rule: preview frames may be overwritten; formal frames require frozen identity, a durable location, and a retryable processing record.
05.2 / FORMAL INSPECTION FLOW
A formal inspection secures identity before optimizing throughput.
The capture callback forks immediately. Preview serves operator visibility; the formal branch builds durable lineage from original image and job to model result and human review.
BACKENDEstablish runtime identity
Confirm the current workpiece, section, and phase
BACKEND → CAPTURESend a complete control snapshot
Include formal-binding permission and frozen context
CAPTUREReceive six-view callbacks
Split preview frames from formal frames at the source
PREVIEW
Deliver the newest view
No business queue or durable record; freshness wins under pressure.
01 / CAPTURE
Persist the original first
Submit location and frozen metadata only after the write succeeds.
02 / INGRESS
Validate, deduplicate, and ACK
ACK means scheduled—not inspected.
03 / WORKER
Durable work and bounded batches
The queue absorbs inference variance; workers form micro-batches.
04 / ALGORITHM
In-process vision pipeline
Read shared images and preserve source coordinates through ROI, anomaly, and defect stages.
05 / BACKEND
Attach lineage and persist
Map results back to frame, camera, section, and model version.
06 / FRONTEND
Emit deltas and review on demand
Lightweight events update live state; REST loads complete evidence.
INVARIANTEven when inference returns late, a result can only update the frame and section frozen at ingress; it must never infer ownership from the browser's current state.
05.3 / KEY SEQUENCE
Fast acknowledgement decouples capture latency from inference latency.
The order and waiting boundaries matter: persist the original before submitting metadata; acknowledge ingress without waiting for the model; send a lightweight delta before loading full evidence on demand.
01Signals
02Backend
03Capture
04Storage / Queue
05Algorithm
06Frontend
A / ESTABLISH RUNTIME CONTEXT
B / ACCEPT FORMAL FRAME
C / INFERENCE AND EVIDENCE LOOP
Only public software contracts are shown. Site addresses, credentials, thresholds, capacity settings, and production performance remain private.
05.4
Communication contract index
01Production signalsRuntime adapterBackend
Create the authoritative workpiece and section context.
02BackendRuntime/control contextCapture
Provide the authoritative identity and gate conditions for formal capture.
03CapturePreview WebSocketFrontend
Deliver latest-only camera and fused views without entering the formal path.
04CaptureJSON ingressBackend
Submit image locations with frozen business context after durable write.
05BackendRedis workers + HTTP batchAlgorithm
Decouple fast acknowledgement from bounded micro-batch inference.
06AlgorithmStructured resultBackend
Return detections that are persisted and mapped to frame and section lineage.
07BackendREST + event WebSocketFrontend
Serve history and detail on demand while pushing lightweight live deltas.
05.5
Detection path, step by step
- 01
Freeze runtime identity
The backend establishes the current workpiece and section; capture freezes that context onto each formal frame.
- 02
Persist before ingress
Capture writes the original image first, then submits its storage location and metadata instead of moving image bytes through the API.
- 03
Acknowledge and queue
The backend validates and quickly acknowledges ingress, placing durable work on Redis so acquisition is not blocked by inference.
- 04
Batch vision inference
Workers group bounded jobs; the algorithm service reads shared images and runs preprocessing, ROI, anomaly, and defect stages.
- 05
Persist evidence and emit deltas
Returned detections are attached to their frozen lineage, stored, and reduced to lightweight realtime events.
- 06
Review and govern
Operators retrieve full evidence on demand; review outcomes feed sample curation and controlled model iteration.
06 / Technical challenges
01
Keep preview out of the formal path
Use latest-only edge streams for operator visibility while reserving durable storage and inference capacity for formally bound frames.
02
Preserve identity across overlap
Freeze workpiece and section context with each formal frame so delayed inference cannot attach evidence to a newer production state.
03
Turn detections into reviewable evidence
Index full-section imagery, map detections back to camera and position, and load detailed evidence on demand instead of flooding the browser.
04
Govern model change
Separate preprocessing, anomaly, and defect model versions, then connect review outcomes and sample governance to controlled iteration.
07 / Outcome
Outcomes bounded by available evidence.
The capabilities below were verified against the internal implementation and maintained engineering documentation. Client identity, site topology, production imagery, throughput, accuracy, alarm rules, and deployment addresses remain private. Every number shown in the public UI is illustrative—not a production result.
01Six-view edge acquisition with fused and focused inspection viewsIMPLEMENTATION VERIFIED
02Durable frame-to-section-to-defect evidence lineageIMPLEMENTATION VERIFIED
03Real-time events, historical retrieval, and image-level human reviewIMPLEMENTATION VERIFIED
04Versioned preprocessing, anomaly, and defect model governanceIMPLEMENTATION VERIFIED