SeasonOCR

SeasonOCR for EasyOCR Models is a .NET OCR library for ONNX models converted from the EasyOCR Python project.

The initial public release focuses on a clean CPU-only API:

• caller-managed long-lived InferenceSession instances
• no complex model path binding inside the library
• EasyOCR-style box detection, recognition, and paragraph grouping
• optional annotated output image and optional debug report

Project Links

• GitHub: SeasonRealms/SeasonOCR
• Models: SeasonEngine/SeasonOCR on Hugging Face
• ONNX export workflow: export-easyocr-onnx.yml

If you prefer to build the ONNX models yourself, clone the repository and run the workflow logic locally.

Origin

• The OCR pipeline is translated and adapted from EasyOCR.
• The intended model inputs are ONNX files converted from EasyOCR Python models.
• Files translated from EasyOCR Python sources include source attribution and an explicit modification notice in the file header.

Features

• EasyOCR-style CRAFT post-processing
• CRNN / CTC recognition
• Beam search by default
• Optional dictionary-aware word beam search
• Perspective rectification for rotated text regions
• Optional rotation TTA, disabled by default
• Structured OCR output with boxes, paragraphs, and summary text
• Optional JPEG annotated image in box-only preview style
• Optional debug report controlled by SeasonOcr.EnableDebugOutput
• Automatic recognizer language discovery from ONNX metadata
• Embedded dictionary loading from recognizer ONNX metadata

Install

dotnet add package SeasonOCR

Quick Start

using System.Text;
using Microsoft.ML.OnnxRuntime;
using SeasonOCR;
using StbImageSharp;

SeasonOcr.EnableDebugOutput = true;

var detectorBytes = File.ReadAllBytes(@"craft_mlt_25k.onnx");
using var detectorSession = new InferenceSession(detectorBytes);

var recognizerBytes = File.ReadAllBytes(@"recognizer_ch_sim.onnx");
using var recognizerSession = new InferenceSession(recognizerBytes);

using var stream = File.OpenRead(@"chinese.jpg");
var image = ImageResult.FromStream(stream, ColorComponents.RedGreenBlue);

var result = SeasonOcr.Detect(
    detectorSession,
    recognizerSession,
    image,
    createAnnotatedImage: true);

Console.WriteLine(result.Summary);

foreach (var paragraph in result.Paragraphs)
{
    Console.WriteLine($"{paragraph.Confidence:P0}: {paragraph.Text}");
}

if (result.AnnotatedImage.Length > 0)
    File.WriteAllBytes(@"output.jpg", result.AnnotatedImage);

if (!string.IsNullOrWhiteSpace(result.DebugReport))
    File.WriteAllText(@"output.debug.txt", result.DebugReport, Encoding.UTF8);

Why This API

This initial API is intentionally session-based:

• You create and reuse InferenceSession instances yourself.
• The library only consumes the prepared detector session, recognizer session, and decoded image.
• This avoids hidden path resolution rules inside the OCR call path.
• It also fits engine-style hosts and service applications that want explicit control over model lifetime.

For the first release, the recommended runtime target is CPU execution.

API Notes

SeasonOcr.Detect(...) current signature:

public static SeasonOcrResult Detect(
    InferenceSession detectorSession,
    InferenceSession recognizerSession,
    ImageResult imageResult,
    bool enableWordBeamSearch = false,
    bool allowRotatedRecognition = false,
    bool createAnnotatedImage = false,
    int beamWidth = 5,
    string? dictionary = null)

Parameter behavior:

• enableWordBeamSearch: enables dictionary-aware decoding when usable dictionary data is available
• allowRotatedRecognition: enables rotation TTA for recognition; default is false
• createAnnotatedImage: when true, result.AnnotatedImage contains a JPEG like output.jpg
• beamWidth: beam width used by beam search
• dictionary: optional in-memory dictionary content; when omitted, embedded model dictionaries are used if present

Multilingual support

• "en", "recognizer_en.onnx", "English"
• "latin", "recognizer_latin.onnx", "Latina"
• "zh-CN", "recognizer_ch_sim.onnx", "中文简体"
• "zh-TW", "recognizer_ch_tra.onnx", "中文繁體"
• "ja", "recognizer_ja.onnx", "日本語"
• "ko", "recognizer_ko.onnx", "한국어"
• "th", "recognizer_th.onnx", "ไทย"
• "ta", "recognizer_ta.onnx", "தமிழ்"
• "te", "recognizer_te.onnx", "తెలుగు"
• "kn", "recognizer_kn.onnx", "ಕನ್ನಡ"
• "ar", "recognizer_ar.onnx", "العربية"
• "ru", "recognizer_ru.onnx", "Русский"
• "hi", "recognizer_hi.onnx", "हिन्दी"
• "bn", "recognizer_bn.onnx", "বাংলা"

Debug Output

Enable debug output before calling the OCR API:

SeasonOcr.EnableDebugOutput = true;

When enabled:

• internal debug logging is emitted through Debug.WriteLine(...)
• SeasonOcrResult.DebugReport is populated
• per-box debug information can be inspected through SeasonOcrResult.DebugBoxes

When disabled:

• debug logging is skipped
• DebugReport stays empty unless explicitly produced by the current flow

Result Contents

SeasonOcrResult can contain:

• Summary: human-readable OCR summary
• Boxes: recognized text boxes kept in the final result
• Paragraphs: grouped paragraph output
• AnnotatedImage: JPEG bytes for the box-only preview image
• DebugBoxes: per-box debug data
• DebugReport: plain-text debug report suitable for saving as output.debug.txt

Models

Recommended model sources:

• Download ready-made ONNX models from Hugging Face
• Or generate them yourself from the repository workflow: export-easyocr-onnx.yml

Recognizer metadata support:

• SeasonOCR reads embedded recognizer charset metadata when available
• SeasonOCR reads recognizer lang_list metadata when available
• Embedded dictionaries such as dict_<lang> are used automatically for word beam search

Output Image

When createAnnotatedImage is true, the library generates a JPEG preview similar to the official EasyOCR examples:

• green detection boxes only
• no confidence text overlay
• suitable for saving directly as output.jpg

Status

Current release scope:

• CPU-first
• stable public OCR API
• session-based integration for application hosts

Planned for later:

• engine-oriented runtime helpers
• GPU provider options
• additional provider-specific optimization layers

License

• The project is distributed under Apache License 2.0.
• See LICENSE and THIRD_PARTY_NOTICES.md for attribution and third-party details.