OCR Text Recognition — OcrBlock¶

OcrBlock recognizes text in any video or image source. Internally it runs the multi-stage PP-OCR pipeline — text detection (DBNet) → optional 0°/180° angle classification → text-line recognition (CRNN/SVTR + CTC decoding) — on each processed frame, raises the recognized regions, and optionally draws them into the video. The block lives in VisioForge.Core.AI (VisioForge.DotNet.Core.AI), implements IVideoProcessingBlock, and has one video Input and one video Output.

graph LR;
    Source-->OcrBlock;
    OcrBlock-->VideoRendererBlock;
    OcrBlock-. OnTextDetected .->App[Your app];

Usage¶

using VisioForge.Core.MediaBlocks;
using VisioForge.Core.MediaBlocks.AI;
using VisioForge.Core.Types.X.AI;

var ocrSettings = new OcrSettings(
    detectionModelPath: "ch_PP-OCRv5_mobile_det.onnx",
    recognitionModelPath: "latin_PP-OCRv5_rec_mobile_infer.onnx",
    characterDictionaryPath: "ppocrv5_latin_dict.txt",
    classificationModelPath: "ch_ppocr_mobile_v2.0_cls_infer.onnx")
{
    Provider = OnnxExecutionProvider.Auto, // CPU / CUDA / DirectML / CoreML
    FramesToSkip = 3,                      // run OCR every 4th frame on live video
    DrawResults = true,                    // burn boxes + text into the frame
};

var ocr = new OcrBlock(ocrSettings);
ocr.OnTextDetected += (sender, e) =>
{
    foreach (var region in e.Regions)
    {
        Console.WriteLine($"{region.Text} ({region.Confidence:P0}) at {region.BoundingBox}");
    }
};

pipeline.Connect(source.Output, ocr.Input);
pipeline.Connect(ocr.Output, videoRenderer.Input);

await pipeline.StartAsync();

Each OcrTextRegion carries the recognized Text, an average Confidence (0..1), an axis-aligned BoundingBox (Rect), and the detection Polygon — the detector's four OcrPoint vertices (top-left, top-right, bottom-right, bottom-left, in source-frame pixels), which may be rotated for slanted text.

Key settings¶

OcrSettings(detectionModelPath, recognitionModelPath, characterDictionaryPath, classificationModelPath = null) sets UseAngleClassifier from whether a classification model path was supplied.

Models and licensing¶

OcrBlock runs third-party ONNX models; the SDK does not ship weights in the NuGet package. The demos ship the Apache-2.0 PP-OCRv5 mobile models (detection, angle classification, Latin recognition) and a Latin dictionary next to the sample executables. PP-OCR supports 100+ languages — download the matching recognition model and dictionary for other languages.

Use with VideoCaptureCoreX and MediaPlayerCoreX¶

OcrBlock implements IVideoProcessingBlock, so it can be registered directly on VideoCaptureCoreX or MediaPlayerCoreX instead of building a manual Media Blocks pipeline:

var ocr = new OcrBlock(ocrSettings);
ocr.OnTextDetected += Ocr_OnTextDetected;

core.Video_Processing_AddBlock(ocr); // before StartAsync (VideoCaptureCoreX)
// player.Video_Processing_AddBlock(ocr); // before OpenAsync/PlayAsync (MediaPlayerCoreX)

await core.StartAsync();

See Using AI blocks with VideoCaptureCoreX and MediaPlayerCoreX for the full processing-block API, insertion order, and lifecycle rules shared by every video AI block.

Use cases¶

• Document and screen capture — recognize text from scanned documents, ID cards, forms, or shared screens in a video conferencing pipeline.
• Retail and warehouse automation — read product labels, barcodes' printed text, or shelf tags from a fixed overhead or handheld camera.
• Industrial inspection — read serial numbers, batch codes, or printed labels on a production line.
• Signage and broadcast monitoring — verify that on-screen text (lower thirds, tickers, digital signage) matches expected content.
• Accessibility tooling — extract on-screen text for text-to-speech or translation pipelines.

For a specific, narrower case — reading vehicle license plates — use the purpose-built License plate recognition (ANPR) block instead of general OCR; it is both more accurate and faster because it runs a plate-specific detector and OCR head rather than scanning the whole frame for any text.

Troubleshooting¶

Frequently Asked Questions¶

What's the difference between OcrBlock and LicensePlateRecognizerBlock?¶

OcrBlock reads arbitrary text anywhere in the frame. LicensePlateRecognizerBlock is a dedicated two-stage pipeline (a plate detector plus a plate-specific OCR head) tuned only for vehicle plates — use it instead of OcrBlock for ANPR/LPR scenarios.

Does OcrBlock support languages other than English?¶

Yes. PP-OCR supports 100+ languages. Point RecognitionModelPath and CharacterDictionaryPath at the recognition model and dictionary for your target language; both must match.

Can I run OCR on a still image instead of a live video stream?¶

Yes — connect a file/image source to OcrBlock.Input in a MediaBlocksPipeline, or feed a single frame through the pipeline; the block processes whatever frames reach its input pad, live or file-based.

Does OcrBlock need a GPU to run in real time?¶

No, but a GPU execution provider (CUDA, DirectML, or CoreML) reduces per-frame latency compared to CPU. For live video, combining FramesToSkip with CPU inference is also a common, GPU-free way to keep OCR from becoming the pipeline bottleneck.

Demos¶

• OCR Text Recognition Demo — WPF Media Blocks pipeline demo.
• OCR Text Recognition MB — the same Media Blocks demo for MAUI.

Dedicated VideoCaptureCoreX/MediaPlayerCoreX OCR demos (Capture OCR X, Capture OCR X WPF, Player OCR X, Player OCR X WPF) are in the SDK's demo set and will be linked here once published to the public samples repository.