Automated visual inspection

Automated visual inspection is the applied form of the pipeline system: detection, classification and segmentation composed into one check that runs continuously, at the point where the product or the parcel actually passes.

It runs on local hardware inside the site network, and reports through structured results rather than a wall of video — counts, labels, scores, boxes and masks, with annotated frames for the people who still need to see.

An inspection station is a graph, and the presets cover the shapes most lines need. Each one can be used as it ships or edited into the station's own version.

• defect_detect_classify — text-prompted defect detection, then classification of every defect region, with a combined result per defect.
• assembly_verification — a whole-part assessment by a vision-language model running in parallel with component detection.
• component_inspection — detect the parts, crop each one, then inspect it individually for damage, orientation or missing sub-parts.
• grounding_dino_sam — detect with text, segment each detection, refine the mask, for checks where geometry matters.

Results arrive as objects rather than pixels: instances with labels, scores and boxes, masks where segmentation ran, and counts per label. Annotations can be rendered over the frame for a human, and the same results exported as JSON or CSV for a system.

Where the decision depends on size, a mask and its box give an area that a rule can compare against a threshold the line already uses, so the acceptance criterion stays in numbers rather than in an inspector's judgement.

• Per-item counts and labels, for counting what passed rather than only flagging what failed.
• Confidence scores that can be filtered, so a review queue is tunable on the floor.
• Annotated images for the operator and structured JSON or CSV for the record.
• Measured areas for defects where size, not presence, decides acceptance.

The framework documents what each pipeline needs, so a station can be sized before it is bought. CPU execution is supported for every one of these checks; it changes the rate, not the capability.

• Surface defect detection and classification: about 200 ms per image on GPU, about 2 seconds on CPU, with a 4 GB GPU as the recommended floor.
• Defect segmentation with SAM and mask refinement: about 500 ms per image on GPU, with 6 GB of VRAM recommended.
• VLM assembly verification: about 1–2 seconds per image, with a GPU expected for practical rates.
• Per-component VLM inspection: about 2–5 seconds per image depending on how many components are inspected.

State the throughput you need first; the pipeline and the hardware follow from that number, not the other way round.

The documented use cases are the places inspection already happens by eye: automotive body panels, printed circuit boards, metal castings, plastic injection moulding, and any process where a camera sees the same product repeatedly.

A second product variant usually does not mean a second project. It means a second set of labels, a second branch in the graph, and a threshold that suits the new part.

• In-line checks that run alongside production rather than in a separate room.
• Per-part inspection where an item passes the camera once and a decision is needed immediately.
• Exception routing, so operators see the cases that need a person.

The practical first step is a folder of frames from the camera you intend to use, good parts included and not only failures. Prompts, labels and thresholds can then be tuned against real images before anything is connected to the line.

Run it offline first and compare the output with the manual check it would replace; the disagreement column is usually more informative than the accuracy figure. Only then decide the rate at which the station runs.