Skip to content

Picture Classification

Use the Picture Classification Rule Type when a field should be validated or derived from an image rather than text. The rule supports image-type validation (e.g., Signature, Fingerprint, Barcode, Face), object and web detections, Safe Search checks, text annotations, and visual similarity via image comparison with thresholds and optional HSV color filtering.

Picture — base options

When to use

  • Confirm an image field contains the expected content (e.g., signature present, barcode/QR present, a face in a selfie box).
  • Enrich records with attributes derived from an image (e.g., face detection flags, best-guess category, web entities).
  • Enforce policy and safety (e.g., block adult/violent content).
  • Validate visual similarity, color presence, or shape features (e.g., brand mark, stamp, background color band).

Open Field Configuration

See Field Rules (Rules Engine) for how to open the field configuration:

  • From the document overlay (supported services), or
  • From the Fields panel on the right sidebar.

Configure the Picture rule

Picture — classification dropdown

Core settings:

  • Rule Type: Picture
  • Default Value: The value to return when nothing is extracted (e.g., Empty Text).
  • Picture Classification: Select the expected image type.

Available classifications:

  • Empty
  • Handwriting
  • Signature
  • Photo
  • Document
  • Barcode
  • QRCode
  • CheckMark
  • Fingerprint
  • Face
  • Scene
  • Other

Panels to enable/configure:

  • Object Detection Options
  • Image Comparison Options
  • Signature Detection Options (visible only when Picture Classification = Signature)
  • Fingerprint Detection Options (visible only when Picture Classification = Fingerprint)
  • Ensure upright orientation (auto-rotate or manual rotate if needed).
  • Prefer original or high-quality scans; avoid heavy compression.
  • Crop to the region of interest if the target occupies a small area.
  • Apply light deskew/denoise/contrast normalization for consistent detection across devices.

Consistent preprocessing

Use the same preprocessing pipeline across environments (scanner, mobile capture) to keep detection thresholds stable.

Object Detection Options

Picture — object detection (Face/Web/Safe/Text)

Toggles:

  • Face Detection

    • Face Detection Options
      • Detect Age — estimate age range.
      • Detect Gender — estimate gender presentation.
      • Detect Emotions — return emotions with likelihood/confidence.
      • Detect Landmarks — return facial landmarks (e.g., eyes, nose bridge).

  • Web Content Detection

    • Web Detection Options
      • Best Guess Category
      • Fully Matching Images
      • Pages with Matching Images
      • Partial Matching Images
      • Visually Similar Images
      • Web Entities

  • Safe Search

    • Safe Search Options
      • Adult
      • Landmarks (include landmark signals where applicable)
      • Allow racy content
      • Violence and Gore

  • Text Annotation

    • Text Annotation Options
      • Full Text — dense captions/paragraph-like text.
      • Text — short tags/captions.

Responsible AI and privacy

Only enable face-related sub-attributes (age/gender/emotions) when legally justified and necessary. Restrict access and log overrides.

Image Comparison Options

Picture — image comparison

Use visual similarity to check whether the image matches a reference pattern/logo or contains specific colors.

Parameters:

  • Accuracy Value
    Minimum similarity threshold (0.00–1.00). Scores below this fail or escalate.

    • The UI “+” control supports adding preset values or multiple reference profiles (where available).

  • Filter Colors
    Enable HSV color filtering before comparison to isolate target colors.

  • Threshold Type
    Method used before/with similarity (e.g., Default, edge/binary variants if present).

  • Threshold Value
    Numeric intensity/cutoff to pair with Threshold Type for noise suppression.

  • HSV Lower Filter / HSV Upper Filter
    One or more HSV ranges (H: 0–360 or 0–179 depending on implementation, S/V: 0–255). Add pairs to capture brand palettes.

Color-first, then shape

When validating logos/stamps with distinctive colors, tune HSV filters first to isolate the signal, then refine Thresholds and Accuracy.

Signature Detection Options (when Picture Classification = Signature)

Picture — signature detection options

These parameters tune the internal signature detector for hand-drawn ink-like strokes and contours. Defaults cover most scanned signatures; adjust if you see false positives/negatives.

  • Adaptive Block Size (default shown in UI example: 11.00)
    Block size for adaptive thresholding; larger values smooth noise in coarse scans.

    • Increase if the background is uneven or if strokes look broken.
    • Decrease if fine strokes disappear.

  • Curvature Coefficient (default: 2.00)
    Emphasizes curved stroke continuity over straight edges.

    • Increase to favor flowing cursive signatures.
    • Decrease if stamps or straight artifacts are being misread as signatures.

  • Contour Count Threshold (default: 35)
    Minimum number of meaningful contours required to accept a signature.

    • Raise to reduce false positives from small scribbles or marks.
    • Lower if genuine signatures are short/initial-like.

  • K Size (default: 9)
    Kernel/window size for smoothing and feature extraction.

    • Larger values smooth noise but can blur detail.
    • Smaller values preserve detail but may admit noise.

  • Threshold (default: 32.00)
    Intensity threshold for binarization/edge extraction.

    • Raise in bright scans to reduce background speckle.
    • Lower if light-ink signatures are getting lost.

Recommended tuning flow:

1) Start with defaults; test 10–20 real samples (pass/fail).
2) Adjust Adaptive Block Size and Threshold until stroke continuity looks stable.
3) Use Contour Count Threshold to eliminate scribble false positives.
4) Fine-tune Curvature Coefficient and K Size for cursive vs printed initials.

Fingerprint Detection Options (when Picture Classification = Fingerprint)

Picture — fingerprint detection options

These parameters target ridge/valley prominence and corner/keypoint salience on fingerprints.

  • Harris Corner Aperture Size (default: 3)
    Aperture for the Sobel operator in the Harris detector.

    • Increase to emphasize broader features; decrease for sharper edges.

  • Harris Corner Block Size (default: 2)
    Neighborhood size considered for corner detection.

    • Larger sizes are more tolerant to noise but may merge fine ridge endings.

  • Harris Corner Kparam (default: 0.02)
    Harris detector sensitivity.

    • Increase to detect more corners (risk: noise).
    • Decrease to be conservative (risk: missing weak minutiae).

  • Key Point Threshold (default: 50.00)
    Minimum score for accepting a keypoint/minutiae candidate.

    • Increase to reduce false positives in low-quality scans.
    • Decrease if high-quality prints still under-detect minutiae.

Recommended tuning flow:

1) Normalize brightness/contrast first (consistent capture).
2) Tune Kparam and Key Point Threshold to get a stable minutiae count on known-good samples.
3) Adjust Aperture/Block Size to balance fine ridge detail vs noise.

Output and formatting

The rule can emit:

  • Booleans: e.g., SignatureDetected, FaceDetected, SafeSearchPassed.
  • Enum/labels: PictureClass, BestGuessCategory.
  • Numerics: DetectionConfidence, SimilarityScore.
  • Objects/metadata: WebEntities list, Face.Emotions, landmark coordinates.

Use “Use Analytics Value/Confidence” (if available) to set the field value and confidence from analytics. Store richer JSON in metadata for audit, while surfacing a simple pass/fail or label to users.

HITL triggers and reviewer guidance

Trigger review when:

  • Expected class not detected (e.g., Signature = false, Fingerprint = false).
  • Similarity/Accuracy score < threshold or within a borderline band (e.g., within ±0.05).
  • Safe Search raises any blocked category, or detectors produce low confidence.
  • Face detections violate policy (e.g., 0 faces or >1 face when exactly one is required).
  • Web Detection suggests mismatches (e.g., “best guess” does not fit policy).

Suggested reviewer note (HITL)

“Verify that the image content matches the expected type (e.g., Signature/Fingerprint). Review detection details and the similarity score against policy. If detections are ambiguous or quality is poor, confirm the image or replace it.”

Examples

  • Signature presence check

    • Picture Classification: Signature
    • Signature Detection Options: defaults
    • Safe Search: On
    • Result: SignatureDetected = true; SafeSearchPassed = true

  • Fingerprint detection with stricter minutiae threshold

    • Picture Classification: Fingerprint
    • Harris Corner Kparam: 0.02 → 0.03; Key Point Threshold: 50 → 65
    • Result: Fewer false positives; escalate if minutiae count < policy threshold.

  • QR code confirmation + text

    • Picture Classification: QRCode
    • Text Annotation: Text = On
    • Result: QRDetected = true; CodeText extracted; escalate if no code or unreadable.

  • Logo verification (color + similarity)

    • Picture Classification: Other
    • Image Comparison: Filter Colors = On; HSV ranges for brand blue; Accuracy Value = 0.85
    • Result: SimilarityScore = 0.90 → pass.

Best practices

  • Enable only the detectors that drive your decision to minimize noise and runtime.
  • Tune thresholds with a representative validation set; document chosen values with examples.
  • Use HSV color filtering for brand/stamp validation; keep ranges tight and test under varied lighting.
  • Store full detection/threshold metadata for audit; display a single status to users.
  • For sensitive attributes (face/age/gender), follow privacy-by-design: disable unless required, restrict access, and log overrides.

Testing checklist

  • [ ] Positive/negative samples for each chosen classification (Signature, Fingerprint, Face, etc.).
  • [ ] Threshold sweeps to set stable Accuracy and Key Point thresholds.
  • [ ] Safe Search scenarios (benign and borderline) in a controlled environment.
  • [ ] HSV ranges tested under different lighting/backgrounds.
  • [ ] Face counts and confidence handling (0, 1, >1).
  • [ ] HITL routing for borderline or conflicting detections.

Troubleshooting

  • Signature false positives (scribbles)
    Raise Contour Count Threshold; increase Curvature Coefficient; adjust Threshold upward to reduce background noise.

  • Signature missing (light ink)
    Lower Threshold; reduce Adaptive Block Size slightly; decrease Contour Count Threshold if genuine signatures are brief.

  • Fingerprint detects too many noisy keypoints
    Increase Key Point Threshold; lower Kparam; slightly increase Block Size.

  • Fingerprint under-detects minutiae
    Decrease Key Point Threshold; increase Kparam; reduce Aperture/Block Size to preserve fine detail.

  • Similarity unstable across lighting
    Normalize exposure/contrast; refine HSV ranges; adjust Threshold Type/Value.

UI reference

  • Base panel (Rule Type, Default Value, Picture Classification)
    Picture — base options

  • Classification dropdown (types)
    Picture — classification dropdown

  • Object Detection Options (Face/Web/Safe/Text)

  • Image Comparison Options (accuracy, thresholds, HSV filters)
    Picture — image comparison

  • Signature Detection Options (visible when classification = Signature)
    Picture — signature detection options

  • Fingerprint Detection Options (visible when classification = Fingerprint)
    Picture — fingerprint detection options