Computer Vision Inspection

Part of the worlds-biggest-software-project initiative.

An AI-native, open-source platform for custom model training and visual quality inspection on the production line.

Computer Vision Inspection is a candidate open-source platform for manufacturers, quality engineers, and inspection service providers who need to automate visual quality control without the cost and rigidity of legacy machine vision systems or the lock-in of cloud-only SaaS. It combines supervised defect detection, unsupervised anomaly detection, and natural-language defect specification in one tool.

Why Computer Vision Inspection?

Industrial machine vision systems (Cognex, Keyence, Omron) deliver reliability but start at USD 5,000–50,000+ per inspection station and require specialist integrators, putting them out of reach for many SMB manufacturers.
Cloud SaaS platforms (Roboflow, Landing AI, AWS Lookout for Vision, Azure Custom Vision) impose recurring costs, vendor lock-in, and round-trip latency that is incompatible with high-speed inline inspection.
Leading open-source detection frameworks such as Ultralytics YOLO are AGPL-3.0, creating IP friction for embedding in commercial products and offering no built-in annotation, anomaly detection, or production feedback tooling.
No current platform productises natural-language defect specification, automated continuous learning from production feedback, or root-cause analysis across multi-station production lines — all identified as underserved areas in the market survey.
Modern vision foundation models have compressed labelled-image requirements from ~50,000 (2023) to 2,000–5,000, making a data-efficient open-source alternative technically viable for the first time.

Key Features

Image Ingestion and Inspection Core

Batch image upload, live camera streams (RTSP/USB), and REST API trigger for on-demand inspection.
Supervised defect detection and classification with fine-tunable object detection models.
Unsupervised anomaly detection trained from normal (good) images only — no labelled defects required.
Pass/fail output with confidence scores, defect coordinates, and per-image result logging.

Labelling and Training Workflow

AI-assisted labelling with bounding box and polygon annotation and model-suggested pre-labels.
Dataset management for image archives, defect type, confidence, timestamp, and pass/fail decision.
Active learning pipeline that routes low-confidence predictions to a human review queue and retrains on accepted corrections.
Synthetic defect image generation from a text description via a diffusion-model backend.

Natural Language and Foundation Models

Natural language defect query at inference time (e.g. "show me parts with surface scratches larger than 3mm").
Designed to leverage open-vocabulary detection approaches (e.g. YOLO-World) and vision foundation models (SAM, DINOv2) for low-shot fine-tuning.

Deployment and Industrial Integration

REST inference API exposing confidence scores and defect coordinates for integration with production systems.
Edge deployment package supporting Docker, NVIDIA Jetson, and Raspberry Pi targets.
OPC UA result output for MES/SCADA integration.
ONNX export pathway so trained models can run across NVIDIA, Intel OpenVINO, and ARM runtimes.

Backlog Capabilities

3D inspection input from point clouds and depth maps.
Multi-station defect correlation with root-cause suggestions across production lines.
Compliance validation workflow for regulated industries (21 CFR Part 11 audit trail, digital signatures).
Native barcode and OCR inspection alongside defect detection.

AI-Native Advantage

The project's AI-native opportunity is grounded in five capabilities that incumbents do not yet productise: natural-language defect specification, foundation-model fine-tuning with 100–500 examples instead of thousands, automated continuous learning from production override feedback, synthetic generation of rare defect images, and multi-station inspection graphs that connect defects back to upstream process steps. Together these compress the data-collection burden that has historically been the primary barrier to deploying custom inspection models.

Tech Stack & Deployment

The platform targets both self-hosted and edge deployment, with Docker containers and packaging for NVIDIA Jetson and ARM SoCs to support latency-sensitive inline inspection. Open standards in scope include ONNX for model interchange, OPC UA for MES/SCADA connectivity, GS1/ISO 15415 for barcode and 2D symbol verification where relevant, and SEMI E187/E188 for semiconductor inspection contexts. Integration is exposed through a REST inference API alongside SDK access patterns familiar to developers using Roboflow, Ultralytics, and the major cloud vision services.

Market Context

The global computer vision market was valued at approximately USD 32 billion in 2026, with manufacturing quality inspection a major sub-segment growing at a double-digit CAGR (Lasting Dynamics, 2026). Incumbent pricing bifurcates between hardware-centric industrial systems at USD 5,000–50,000+ per station (Cognex, Keyence) and SaaS platforms (Roboflow from USD 249/mo; AWS Lookout at USD 4–8 per 1,000 images; Landing AI enterprise-only). Primary buyers are QA engineers and operations directors at electronics, automotive, food, and pharmaceutical manufacturers, plus third-party inspection service providers seeking to scale without adding headcount.

Project Status

This project is in the research and specification phase.
Contributions, feedback, and domain expertise are welcome.

Contributing

We welcome contributions from developers, domain experts, and potential users. See CONTRIBUTING.md for guidelines.

Important: All contributions must be your own original work or clearly attributed open-source material with a compatible licence. Copyright infringement and licence violations will not be tolerated and will result in immediate removal of the offending contribution. If you are unsure whether a piece of code, text, or other material is safe to contribute, open an issue and ask before submitting.

Licence

Licence to be determined. See discussion for context.