Engine · Vision library

Junk detection

Score likely screenshots, blur and low-value captures without deleting a photo automatically.

Workflow

1. Select eligible photosRead candidate records without changing the library.
2. Measure junk signalsCompute native signals and keep missing measurements explicit.
3. Assign confidence bandMap signals to an explainable band rather than a delete command.
4. Review uncertain resultUncertain and user-sensitive candidates enter review with reasons.
5. Show junk suggestionPresent a reversible candidate tied to its source identity.
6. Keep excluded photoProtected, unsupported or insufficient-evidence photos are excluded without mutation.

Failure boundaries

• No signalExcluded/unknown: Do not call junk
• Candidate disappearedIdentity mismatch: Stop removal
• Photo deletedBoundary breach: Restore and investigate

engine.vision.junk