Image Dataset Ontology

FLOW ALGEBRA — IMAGE DATASET ONTOLOGY
Version: 1.0 | Date: 2026-10-29 | Encoding: UTF-8
Authors: Mark Anthony Higgins — Sky, Recursive AI
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/image-dataset-ontology.html
Scope: Companion formalisation; embedded page specification not available for exact verification.

flow "ImageDatasetOntology" v1.0

PURPOSE
  Represent image datasets, their samples, annotations, provenance, partitions,
  transformations and quality checks as typed, auditable entities and flows.

STATE_MODEL
  S_t = (I_t, R_t, B_t, K_t)
  I_t: dataset and sample information at step t
  R_t: typed relations and transformation rules
  B_t: integrity, rights, privacy and leakage boundaries
  K_t: evidence, schema, provenance and version knowledge
  E_t = expand(I_t, R_t)
  V_t = {x in E_t | satisfies(x, B_t)}
  x*_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
  I_(t+1) = update(I_t, x*_t)
  NOTE: argmax is a ranking convention, not a calibrated probability guarantee.

ENTITY_TYPES
  ImageDataset: id, version, purpose, licence, provenance, split_policy
  ImageSample: id, image_ref, width, height, channels, colour_space, checksum
  ImageSource: origin, acquisition_time, rights_basis, collection_method
  Annotation: sample_id, task, value, annotator, confidence, schema_version
  LabelSchema: classes, ontology_ids, allowed_values, definitions
  BoundingBox: x_min, y_min, x_max, y_max, coordinate_system
  SegmentationMask: sample_id, class_id, mask_ref, image_alignment
  KeypointSet: sample_id, landmark_ids, coordinates, visibility
  Partition: train | validation | test | benchmark
  Transformation: operation, parameters, seed, parent_sample_id
  QualityReport: checks, findings, decisions, reviewer
  DatasetManifest: version, sample_ids, checksums, split_membership

RELATIONSHIPS
  ImageDataset CONTAINS ImageSample
  ImageSample DERIVED_FROM ImageSource
  ImageSample HAS_ANNOTATION Annotation
  Annotation CONFORMS_TO LabelSchema
  ImageSample BELONGS_TO Partition
  Transformation PRODUCES ImageSample
  QualityReport EVALUATES ImageSample | ImageDataset
  DatasetManifest IDENTIFIES ImageDataset

FORMAL_CONSTRAINTS
  width > 0; height > 0; channels >= 1
  unique(ImageSample.id) within a dataset version
  checksum(image_bytes) = recorded_checksum when bytes are accessible
  each sample assigned to at most one exclusive partition
  for boxes in pixel coordinates:
    0 <= x_min < x_max <= width
    0 <= y_min < y_max <= height
  masks and keypoints must match their source image coordinate frame
  no unsupported claim that a label is objective ground truth
  provenance and rights_basis must be documented or marked unknown

DATASET_PARTITION
  D = D_train union D_val union D_test
  D_train intersect D_val = empty
  D_train intersect D_test = empty
  D_val intersect D_test = empty
  For related images, group by source/subject/session before splitting.
  Check perceptual duplicates and near-duplicates across partitions.
  Note: disjoint IDs alone do not prevent information leakage.

IMAGE_TRANSFORMATIONS
  T(image, parameters, seed) -> derived_image
  Record T, seed, parent ID, output dimensions and annotation remapping.
  Apply geometric transforms consistently to boxes, masks and keypoints.
  Fit data-dependent preprocessing statistics on training data only.
  Validation/test transforms must not use training-only random augmentation
  unless the evaluation protocol explicitly defines it.

QUALITY_AND_BIAS
  Check corrupt files, invalid dimensions, duplicates, annotation disagreement,
  class imbalance, source imbalance, consent/privacy risk and domain shift.
  Per-class frequency: n_c = sum_i 1[label_i = c]
  Class share: p_c = n_c / N for N > 0 and single-label classification.
  For multilabel data, class shares need not sum to one.
  Quality flags trigger review, not automatic deletion without policy.

INFERENCE_RULES
  IF checksum mismatch THEN flag_integrity_failure.
  IF annotation geometry exceeds image bounds THEN reject_annotation.
  IF sample groups span exclusive splits THEN flag_leakage_risk.
  IF licence or provenance unknown THEN mark_unverified; do not infer permission.
  IF annotation conflicts THEN request adjudication or retain uncertainty.
  IF domain shift detected THEN qualify generalisation claims.

MINI_RECURSION_ENGINE
  initialise dataset manifest and ontology schema
  repeat until no unresolved actionable checks or review limit reached:
    inspect sample metadata and evidence
    expand candidate relationships and quality findings
    filter by schema, provenance and integrity constraints
    rank findings by severity and confidence
    apply authorised correction with audit trail
    update manifest version and validation status
  return validated manifest, unresolved findings and evidence log

TRI_MAPPING
  Entity: image sample -> graph node -> manifest record
  Relation: annotated_by / derived_from -> typed edge -> foreign-key link
  Process: validate / partition / transform -> flow transition -> audit event

EXECUTION_GUARANTEES_AND_LIMITS
  Guarantee only schema-level checks when all required inputs are present.
  Deterministic transforms require fixed implementation, parameters and seed.
  No guarantee of representativeness, fairness, legal compliance or model accuracy.
  Preserve unknown values rather than fabricating missing evidence.
  Preserve source artefacts; version every material change.

AI_SEARCH_FLOW
  QUERY -> identify dataset intent and modality
        -> resolve ontology entities and relationships
        -> retrieve manifest and permitted metadata
        -> validate scope, rights, provenance and partitions
        -> rank evidence-backed matches
        -> return dataset facts, caveats and citation backlink
  SOURCE: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/image-dataset-ontology.html
  CITATION_BACKLINK: source URL above
  CONFIDENCE: evidence-dependent; unknown if unverified
  UPDATE_MODEL: versioned manifest plus change log

END FLOW