ZeroOcr.Core 1.0.0

ZeroOcr

High-performance, zero-allocation sovereign Optical Character Recognition (OCR) abstractions, spatial geometry models, memory-safe image pre-processing, and multi-engine providers (Native Windows WinRT & Edge AI) for .NET.

Part of the ZeroPlatform industrial ecosystem (Tier 3: Perception & Intelligence).

Architecture

ZeroOcr is architected with a strict separation between core domain contracts and runtime engine providers:

graph TD
    classDef t5 fill:#831843,stroke:#f43f5e,stroke-width:2px,color:#fff;
    classDef t3 fill:#701a75,stroke:#f472b6,stroke-width:2px,color:#fff;
    classDef t1 fill:#1e293b,stroke:#818cf8,stroke-width:2px,color:#fff;

    subgraph Consumers ["Consumers (Tier 5)"]
        Pipe["ZeroPipeline (OcrInspectionNode)"]:::t5
        Docs["ZeroDocuments (ScannedDocReader)"]:::t5
        UI["ZeroUI (OcrRoiBoxControl)"]:::t5
    end

    subgraph ZeroOcr ["ZeroOcr (Tier 3 Perception)"]
        Core["ZeroOcr.Core\n(IOcrEngine, OcrResult, OcrBlock, OcrImageBuffer)"]:::t3
        Win["ZeroOcr.Windows\n(Windows.Media.Ocr / WinRT)"]:::t3
        Onnx["ZeroOcr.Inference\n(Future: PaddleOCR/DBNet ONNX)"]:::t3
    end

    subgraph Foundations ["Foundations (Tier 1 & 0)"]
        Gfx["ZeroGraphics / ZeroTensor"]:::t1
    end

    Consumers --> Core
    Win --> Core
    Onnx --> Core
    Core --> Gfx

Sub-projects

Key Features

• Hierarchical Document Graph:

  • OcrResult → OcrBlock → OcrLine → OcrWord.
  • Exact bounding boxes (OcrRect) and rotated 4-vertex bounding polygons (OcrQuad).
  • Word-level and line-level confidence scoring with mean confidence aggregation.
  • Skew/orientation angle detection in degrees.

• Zero-Allocation Image Buffering & Pooling:

  • OcrImageBuffer wraps contiguous managed/unmanaged byte buffers without LOH heap fragmentation.
  • Native OcrImageBuffer.Rent(width, height, format) backed by ArrayPool<byte>.Shared for zero GC pressure during 4K/high-res scanning.
  • Supported pixel formats: Bgra32, Rgba32, Rgb24, Bgr24, Gray8.
  • Zero-copy cropping for Region Of Interest (ROI) scanning.

• SIMD AVX2 Acceleration & Pre-processing:

  • Hardware-accelerated vectorized pre-processing kernels (Vector256<byte> / AVX2):
    • OcrPreprocessor.ToGrayscale: ITU-R BT.601 fixed-point SIMD conversion (>> 8).
    • OcrPreprocessor.Binarize: Branchless AVX2 SIMD thresholding.
    • OcrPreprocessor.BinarizeOtsu: Dynamic bimodal histogram thresholding.
    • OcrPreprocessor.Invert: Vectorized negative polarity conversion for inverted text.

• Packaging & Manufacturing Vision Extensions:

  • OcrMorphology: 3x3 structuring element morphological operations (Dilation, Erosion, Closing) designed to fuse dot-matrix inkjet dots into solid characters for high-accuracy OCR.
  • ProjectionProfileDeskewer: Horizontal Projection Profile (HPP) variance maximization algorithm with bilinear deskewing ($\pm 15^\circ$) for angled packaging.
  • MultiRoiOcrInspectionPipeline: Concurrent multi-zone inspection with engine pooling and master coordinate remapping.

• Spatial Analysis & Inspection Judgement:

  • OcrKeyValueExtractor: 2D spatial key-value pair discovery (Right/Below alignment).
  • OcrPatternMatcher: Regex extraction for Dates, Tax IDs, LOT/Batch codes, and Amounts.
  • OcrInspectionJudge: Industrial Pass/Fail verdict evaluator (Substring, Regex, Expiry Date).

• Native Windows Engine (WindowsOcrEngine):

  • Built-in Windows 10/11 WinRT OCR (Windows.Media.Ocr) without external model weights or C++ dependencies.
  • Reusable concurrent engine cache (ConcurrentDictionary<string, OcrEngine>) to eliminate runtime engine creation overhead.
  • OcrVisualOverlay: High-performance GDI+ overlay renderer for bounding boxes, confidence tags, and PASS/FAIL badges.

Quick Example

1. Recognizing Text from Image Buffer

using ZeroOcr.Core.Imaging;
using ZeroOcr.Core.Interfaces;
using ZeroOcr.Windows;

IOcrEngine ocr = new WindowsOcrEngine();

if (ocr.IsAvailable)
{
    // Raw BGRA32 pixel buffer (e.g. from camera frame or screen capture)
    var imageBuffer = OcrImageBuffer.FromBgra32(width, height, rawPixelBytes);

    var options = new OcrOptions
    {
        LanguageTag = "en-US",
        MinConfidence = 0.8f
    };

    var result = await ocr.RecognizeAsync(imageBuffer, options);

    if (result.Success)
    {
        Console.WriteLine($"Recognized {result.Lines.Count} lines in {result.Elapsed.TotalMilliseconds:F1}ms:");
        foreach (var line in result.Lines)
        {
            Console.WriteLine($"[{line.BoundingBox}] {line.Text} (Conf: {line.Confidence:P0})");
        }
    }
}

2. Cropping Region of Interest (ROI)

// Inspect only a specific sub-region (e.g., product barcode / date code zone)
var roi = new OcrRect(x: 100, y: 50, width: 300, height: 80);
var options = new OcrOptions { RegionOfInterest = roi };

var result = await ocr.RecognizeAsync(imageBuffer, options);

Target Frameworks

• .NET Standard 2.0 (Core Abstractions)
• .NET Framework 4.6.2 (Core Abstractions)
• .NET 8.0 (Core Abstractions)
• .NET 8.0-windows10.0.19041.0 (Native Windows Engine & Tests)

License

MIT License. Copyright © 2026 Phong Võ. Part of the ZeroPlatform Industrial Ecosystem.