VISUAL INSPECTION · LONGITUDINAL MEMORY

The inspection that remembers what it saw.

Upload a capture. Afterimage checks its quality, anchors it to the stored baseline, measures the change and stops when a person must decide.

HOW THE LOOP BRANCHES

Four captures, four different answers.

An illustration of the path each capture takes. Try them in the app: the first photo sets the reference, and each result links the next.

1 · the reference photo

nothing to compare against yet, so it becomes the baseline

  1. 1 assess_quality done
  2. 2 align_to_baseline skipped
  3. 3 diff_against_memory skipped
  4. 4 crop_and_rescan skipped
  5. 5 classify_severity skipped
result first_baseline
18 / 29 real photographs most scenarios use licensed field imagery
0.8621 branch accuracy the final action selected by deterministic policy
0.8542 defect macro F1 across crack, delamination, hotspot and soiling
0.7875 mean IoU localisation over scenarios with a detected region

HOW IT WORKS

The model describes. The policy decides.

Six perception tools produce measurements; one policy module owns every threshold.

01

Capture

Reject unusable images before they can contaminate memory.

02

Compare

Align the same asset, isolate the changed region and measure it.

03

Act

Record mild changes, refuse uncertain matches or ask a person to approve.

MEMORY

A baseline is superseded, never overwritten.

Every asset keeps its inspections, the reference in force and the exact trace behind each result.

THE SYSTEM

One inspectable pipeline, end to end.

OpenCV 5 perceptionFastAPI applicationDynamoDB memoryS3 images and tracesarm64 Lambda delivery

Honest limits

  • 29 evaluation scenarios are still a small sample.
  • One global frame-coverage threshold does not generalise.
  • A severe finding still requires a human decision.

Inspect the next capture. Keep the reason.

Run the real pipeline with the bundled examples or your own image.

Try the four samples