flutter_ocr_native 0.3.3

Flutter plugin for on-device OCR using ML Kit (Android), Vision (iOS/macOS), WinRT (Windows), Tesseract (Linux). Document parsing, validation, and masking.

0.3.3

Improvements

• Old & new document format support — all parsers now handle both legacy and current Indian document layouts:

  • Aadhaar: new bilingual format (नाम / Name, पिता का नाम / Father's Name, जन्म की तारीख / DOB, पता / Address) detected alongside old English-only labels. DOB value on next line supported
  • PAN: bilingual layout already supported from 0.3.1 — no change
  • Driving License: new smart card format (COV, DOI, NT VALIDITY, NON-TRANSPORT validity dates on separate lines) + improved DL number regex covering all Indian state formats including 6-digit serial and compact no-separator variants
  • Voter ID: new format (Name:, Father's Name:, Husband's Name:, Sex:, Age:) parsed alongside old ELECTOR'S NAME / FATHER'S NAME format. Age field converted to approximate birth year when full DOB is absent. M/F single-char gender values normalised to Male/Female
  • Passport: MRZ (Machine Readable Zone) parsing added — P<IND<SURNAME<<GIVEN and digit line parsed to extract surname, given name, nationality, DOB, gender, expiry and passport number directly from MRZ. MRZ date YYMMDD converted to DD/MM/YYYY with correct century. Old labeled-field booklet format unchanged
  • Cheque: new CTS printed format — A/C No., Account No., Acc No., Account Number label patterns parsed before falling back to raw digit scanning. Account number on next line after label also handled

• Smart handwriting detection — per-document-type policy prevents false rejections:

  • Passport and Cheque and Driving License use allowMixed policy — only rejected if zero printed keywords found (truly blank handwritten paper). Signatures on passports and handwritten amounts/payees on cheques no longer cause HandwrittenTextException
  • Aadhaar, PAN, Voter ID keep strict rejectIfHandwritten policy — fully handwritten documents correctly rejected
  • docType hint parameter added to all OcrReader read methods (readFromPath, readFromBytes, readFromFile, readFromPdf, readFromPdfFile) — pass known type to skip auto-detection and apply correct policy immediately

• Document type detector — passport MRZ scoring added: P<IND< pattern scores +5, MRZ digit line pattern scores +5 — ensures MRP passports detected even without the word "PASSPORT" in OCR text

• API surface reduced — package now exports only what app developers need:

  • Removed from public API: OcrPlatformInterface, OcrMethodChannel (internal channel wiring), AadhaarDetails, ChequeDetails, DrivingLicenseDetails, PassportDetails, VoterIdDetails (use DocumentDetails instead), OcrException base class (catch EmptyImageException / HandwrittenTextException directly), OcrDetailsCard, VoterIdDetailsCard (build UI from DocumentDetails.toDisplayMap()), HandwritingPolicy (internal enum)
  • All removed classes still exist in the package source and are used internally — only their re-export from the barrel is removed

• DocumentDetails internal refactor — removed redundant _parse() indirection; fromResultSync now owns the dispatch switch directly. _hasPhoto helper inlined. fromText simplified to single expression

Bug Fixes

• Fixed NOT_INITIALIZED crash on re-upload (Android, Windows, Linux) — scanning a second image after the first scan threw PlatformException(NOT_INITIALIZED, Recognizer not initialized)
  • Android: processImage() now lazily recreates TextRecognizer if null — seamless re-use after dispose()
  • Windows: RecognizeFromBytes() now lazily recreates OcrEngine if null — recovers from failed init or unexpected null state
  • Linux: recognize() now lazily reinitializes TessBaseAPI if null — recovers from failed init or dispose
  • iOS/macOS: unaffected — Vision framework creates a fresh VNRecognizeTextRequest per call with no persistent state

0.3.1

Bug Fixes

• Critical PAN detection fix (Android, iOS, macOS) — bilingual PAN cards (Hindi + English) with no-vowel PAN numbers like BWPPM8548F were silently dropped by the isEnglish() filter

  • Root cause: vowel check letters.count >= 4 && !hasVowel incorrectly rejected valid PAN tokens that contain no vowels (e.g. BWPPM, BWPPMF)
  • Fix applied to all three Vision/ML Kit platforms: Android (OcrPlugin.kt), iOS (OcrPlugin.swift), macOS (OcrPlugin.swift)
  • Any token containing digits (IDs, codes, PAN numbers, dates) now always passes through the English filter
  • Vowel check now only applies to pure-letter strings of 5+ characters (rejects Hindi/non-Latin words)
  • Threshold raised from 4 → 5 letters to avoid rejecting short English abbreviations (DEPT, GOVT, CARD)
  • Windows (WinRT) and Linux (Tesseract) were unaffected — their isEnglish only checks for [A-Za-z0-9] presence with no vowel filter
  • PAN cards with vowels in the number (e.g. AXEPN1010E) were unaffected on all platforms

• PAN extraction — space-tolerant matching — extractPAN() now handles OCR splitting PAN into BWPPM 8548F (space between letter-block and digit-block)

  • Added Strategy 2: ([A-Z]{5})\s+(\d{4})\s*([A-Z]) pattern
  • Added Strategy 3: OCR misread correction (O→0, I→1, S→5) with space tolerance
  • Added Strategy 4: relaxed 4th-char fallback for OCR-misread holder type character

• PAN name/father parsing — bilingual layout — _fromPan() now correctly extracts name and father name from bilingual cards where Hindi labels are stripped to / Name, / Father's Name

  • Added _extractNameValue() helper — checks text after / on same line first, then next line
  • Handles both old format (name on next line after label) and new bilingual format (label + value on same line separated by /)

• Raw OCR text copy button — example app now shows a copy icon on the Raw OCR Text card for easy debugging

• Native PDF to image conversion — render PDF pages to images without any third-party Dart package

  • Android: PdfRenderer (API 21+)
  • iOS: CGPDFDocument (CoreGraphics)
  • macOS: CGPDFDocument (CoreGraphics)
  • Windows: Windows.Data.Pdf (WinRT)
  • Linux: not supported (returns null gracefully)

• Added OcrDocumentSaver.renderPdfPage() — renders a single PDF page to JPEG bytes

• Added OcrDocumentSaver.getPdfPageCount() — returns total pages in a PDF

• Added OcrDocumentSaver.renderAllPdfPages() — renders all pages to a list of image bytes

• Added OcrReader.readFromPdf() — single-call PDF to OCR text

• Added OcrReader.readFromPdfFile() — OCR directly from a PDF file

• Capture instructions widget — OcrCaptureInstructions shows best practices before scan/upload

  • showAsBottomSheet() — modal bottom sheet with tips
  • showAsDialog() — dialog with tips
  • Inline widget mode — embed directly in your UI
  • Customizable instructions via OcrInstruction model

• Enhanced PAN card detection — OCR misread correction (O→0, I→1, S→5), uppercase fallback, name/DOB/father name parsing

• OOM crash fix (Android) — correctOrientation now uses 1024px downscaled test bitmaps for rotation detection instead of full-size images

• OOM crash fix (all platforms) — PDF rendering capped at 3000px max dimension

• Background thread rendering — PDF rendering runs off main thread on Android, iOS, and macOS

• Fixed FlutterImageDecoderImplDefault crash — PDF bytes no longer passed to image decoder; detected and rendered to JPEG first

• Fixed Windows compile error — added Windows.UI.h include, fixed ColorHelper usage with direct struct

• Fixed Windows stream crash — DataWriter.DetachStream() prevents premature stream closure

• Linux graceful fallback — renderPdfPage returns null, getPdfPageCount returns 0 (no crash)

• Example app now supports PDF file picking and processing

• Example app shows capture instructions before camera/gallery

• Added _isPdf() detection in example flow

• File picker now accepts PDF alongside image formats

• correctOrientation skips images >10MB to prevent OOM

• Windows compile fix — resolves all 3 MSVC build errors:

  • Added #define NOMINMAX and (std::max)(...) parenthesization to prevent error C2589 from Windows max macro
  • Replaced broad using namespace WinRT imports with namespace aliases (ocr::, streams::, imaging::, pdf::) to fix error C2872: 'IUnknown' ambiguous symbol
  • Added #include <winrt/Windows.Globalization.h> to fix linker error for Windows::Globalization::Language

0.2.1

• Smart auto-orientation — only rotates if original is unreadable; keeps already-readable images untouched
• Cheque MICR fix — MICR line (special font at bottom) no longer triggers handwriting rejection
• Cheque parser rewrite — case-insensitive IFSC extraction, spaced account numbers, 30+ bank name patterns, address extraction from branch lines
• Added address field to ChequeDetails and DocumentDetails for cheques
• Improved detectPrinted on all platforms — excludes low-confidence numeric elements at image bottom from scoring

0.2.0

• Auto-orientation correction — detects correct image rotation using OCR confidence scoring across all 4 rotations (0°, 90°, 180°, 270°). Works even without EXIF data
• Image cropper with rotate — added rotate button (90° clockwise) to crop UI
• Document type auto-detection — scores OCR text against keywords/patterns for Aadhaar, PAN, Passport, Driving License, Voter ID, Cheque
• Unified DocumentDetails model — single fromResult() API handles all doc types, face extraction, validation, and toDisplayMap(maskAadhaar: true)
• Face extraction — ML Kit Face Detection (Android), Vision (iOS/macOS). Returns cropped face bytes
• Document parsers — PassportDetails, DrivingLicenseDetails, VoterIdDetails, ChequeDetails
• Extended validation — Passport, Driving License, Voter ID (EPIC), IFSC, Account number
• Added OcrDocumentSaver.correctOrientation() — auto-corrects image orientation before display
• Added OcrDocumentSaver.extractFace() and extractFaceFromPath()
• Added OcrDocumentSaver.isFaceExtractionSupported getter
• Added result.docType and result.docTypeLabel getters on OcrResult
• Added DocumentTypeDetector.detect(), .label(), .icon() static methods
• Added OcrDocumentViewer support for originalBytes parameter
• Fixed Voter ID extraction — NFKD normalization for Unicode lookalike characters
• Fixed false Voter ID matches from date strings (e.g., "biRTH 15/07/199x")
• Added macOS platform support — Apple Vision framework
• Added Windows platform support — WinRT OCR engine + GDI+
• Added Linux platform support — Tesseract OCR + Leptonica
• Added DocumentNumberValidator — validates Aadhaar (Verhoeff checksum) and PAN (format + holder type)
• Added AadhaarDetails model — parses OCR text into structured fields
• Added OcrDetailsCard widget
• Fixed Aadhaar masking regex — requires separators, won't mask pincodes
• Fixed validation working on masked text — validates against raw text internally

0.0.7

• Added OcrImageFormat enum — configurable output format (JPEG or PNG)
• Added imageQuality parameter to all save/download methods — JPEG compression 1-100
• Added format parameter to all save/download methods — choose JPEG or PNG output
• downloadFromPath auto-detects output format from original file extension (.png → PNG, others → JPEG)
• Added compressImage() standalone method — compress any image bytes natively
• Accepts any input image format (JPEG, PNG, WEBP, BMP, GIF, HEIC, TIFF) — decoded natively
• Native compression via Android Bitmap.compress and iOS jpegData/pngData
• Watermark is now fully optional — omit or pass null to skip

0.0.6

• Lowered SDK constraint to >=3.2.4 <4.0.0 (Flutter 3.19.0+) for broader compatibility
• Fixed Color.toARGB32() not available on older Dart versions — replaced with version-safe _colorToArgb() helper using .a/.r/.g/.b float API
• Zero deprecation warnings on all Dart 3.x versions (3.2.4 through 3.11+)

0.0.5

• Updated README with complete usage documentation for all features
• Added examples for Basic OCR, Validation & Aadhaar Masking, Document Viewer, Download with Watermark, Custom Validator, and Runtime Toggle
• Added full architecture tree in README covering all source files
• Added Supported Platforms table and Flutter Compatibility section
• Updated Getting Started version to ^0.0.4

0.0.4

• Fixed watermark not appearing in downloaded images
• Moved watermark rendering from dart:ui Canvas to native platform (Android Canvas / iOS CoreGraphics) for reliable text rendering
• Added burnWatermark native method channel — watermark is now burned into images on the native side
• Auto-scaled watermark font size to 3% of image width (minimum 36px) — always readable regardless of image resolution
• Bold watermark text with 1.5x line height for better readability
• Added downloadBytes() method to OcrDocumentSaver for saving raw bytes directly
• Added path_provider as plugin dependency — platform-specific download paths handled internally
• Removed path_provider dependency from example app — package handles it
• Simplified OcrDocumentViewer save — uses native burnWatermark instead of unreliable RepaintBoundary capture

0.0.3

• Added OcrWatermark model — configurable watermark with key-value lines (Lead ID, Lat, Long, Agent, Date, etc.), customizable text color, background color, font size, and padding
• Added OcrDocumentViewer widget — full-screen document viewer with pinch-to-zoom (0.5x–5x), watermark overlay below image, configurable save button, and OcrDocumentViewer.show() for one-liner navigation
• Added OcrDocumentSaver utility — saves masked/original image to file with watermark burned into the image using Canvas, supports save from file path or raw bytes
• Viewer save captures the watermark in the exported image via RepaintBoundary
• Updated example app with View and Download buttons using the new package utilities

0.0.2

• Renamed package from ocr to flutter_ocr_native
• Lowered SDK constraint to support Flutter 3.27.1+
• Fixed Aadhaar image masking for different card positions and orientations
• Improved handwriting detection using ML Kit confidence signals
• Added maskedImageBytes — image with Aadhaar digits blacked out
• Added hasAadhaar getter on OcrResult

0.0.1

• Initial release
• On-device OCR using ML Kit (Android) and Vision framework (iOS)
• English-only text extraction — non-Latin scripts auto-filtered
• Structured results: blocks → lines → elements with bounding boxes & confidence
• Aadhaar number masking (text + image) — configurable
• Handwriting detection — rejects non-printed documents
• Empty/blank image detection