0cve0/OpenMLKitOCR

1---2license: apache-2.03tags:4- ocr5- text-detection6- text-recognition7- tflite8- mlkit9- google-mlkit10- on-device11- image-to-text12pipeline_tag: image-to-text13language:14- en15- ru16- zh17- ja18- ko19- ar20- he21- bn22- gu23- kn24- ml25- ta26- te27- ka28- vi29---30 31# 🏛️ OpenMLkit OCR Models32 33This repository hosts a collection of highly optimized, lightweight, on-device OCR (Optical Character Recognition) models extracted from Google ML Kit APK components. These models are designed to run fully offline and are fully compatible with [OpenMLkitOCR](https://github.com/0cve0/OpenMLkitOCR), a lightweight offline OCR engine for Python.34 35The repository includes:36* **Text Detection Model**: A Region Proposal Network (RPN) architecture (`rpn_detector.tflite`) that identifies text bounding boxes in image tiles.37* **Text Recognition Models**: Lightweight CRNN + CTC model pipelines for 15+ different languages and scripts.38 39---40 41## 🏛️ Supported Languages and Scripts42 43Below is the registry of the available models, vocabulary maps, and language model priors stored in this repository:44 45| Code | Script / Language | Recognizer Model | Label Map File | Language Model / Priors |46| :--- | :--- | :--- | :--- | :--- |47| `detector` | **Text Detection (All)** | [rpn_detector.tflite](./detector/rpn_detector.tflite) | — | — |48| `en` | **Latin / English** | [line_recognizer.fb](./en/line_recognizer.fb) | [LabelMap.pb](./en/LabelMap.pb) | — |49| `ru` | **Cyrillic / Russian** | [recognizer_cyrl.tflite](./ru/recognizer_cyrl.tflite) | [LabelMap_cyrl.pb](./ru/LabelMap_cyrl.pb) | FST LM + Priors |50| `zh` | **Chinese / Han (Hani)** | [recognizer_hani.tflite](./zh/recognizer_hani.tflite) | [recognizer_hani_label_map.pb](./zh/recognizer_hani_label_map.pb) | FST LM + Priors |51| `ja` | **Japanese (Jpan)** | [recognizer_jpan.tflite](./ja/recognizer_jpan.tflite) | [recognizer_jpan_label_map.pb](./ja/recognizer_jpan_label_map.pb) | FST LM + Priors |52| `ko` | **Korean (Kore)** | [recognizer_kore.tflite](./ko/recognizer_kore.tflite) | [recognizer_kore_label_map.pb](./ko/recognizer_kore_label_map.pb) | FST LM + Priors |53| `ar` | **Arabic (Arab)** | [recognizer_arab_retrained.tflite](./ar/recognizer_arab_retrained.tflite) | [recognizer_arab_label_map.pb](./ar/recognizer_arab_label_map.pb) | FST LM + Priors |54| `he` | **Hebrew (Hebr)** | [hebr.tflite](./he/hebr.tflite) | [hebr_label_map.pb](./he/hebr_label_map.pb) | Priors |55| `ka` | **Georgian (Geor)** | [geor.tflite](./ka/geor.tflite) | [geor_label_map.pb](./ka/geor_label_map.pb) | Priors |56| `bn` | **Bengali & Devanagari (Bede)** | [bede.tflite](./bn/bede.tflite) | [bede_label_map.pb](./bn/bede_label_map.pb) | Priors |57| `gu` | **Gujarati (Gujr)** | [gocr_tflite_recognizer_gujr.tflite](./gu/gocr_tflite_recognizer_gujr.tflite) | [gocr_tflite_recognizer_gujr_label_map.pb](./gu/gocr_tflite_recognizer_gujr_label_map.pb) | Priors |58| `kn` | **Kannada (Knda)** | [recognizer_knda.tflite](./kn/recognizer_knda.tflite) | [recognizer_knda_label_map.pb](./kn/recognizer_knda_label_map.pb) | FST LM + Priors |59| `ml` | **Malayalam (Mlym)** | [recognizer_mlym.tflite](./ml/recognizer_mlym.tflite) | [recognizer_mlym_label_map.pb](./ml/recognizer_mlym_label_map.pb) | FST LM + Priors |60| `ta` | **Tamil (Taml)** | [recognizer_taml.tflite](./ta/recognizer_taml.tflite) | [recognizer_taml_label_map.pb](./ta/recognizer_taml_label_map.pb) | FST LM + Priors |61| `te` | **Telugu (Telu)** | [recognizer_telu.tflite](./te/recognizer_telu.tflite) | [recognizer_telu_label_map.pb](./te/recognizer_telu_label_map.pb) | FST LM + Priors |62| `vi` | **Vietnamese / Latin** | [gocr_tflite_recognizer_latn_vi.tflite](./vi/gocr_tflite_recognizer_latn_vi.tflite) | [gocr_tflite_recognizer_latn_vi_label_map.pb](./vi/gocr_tflite_recognizer_latn_vi_label_map.pb) | Priors |63 64---65 66## 📂 File Types Explained67 681. **`*.tflite` / `*.fb` (Neural Network weights)**: 69   * `detector/rpn_detector.tflite` is a Convolutional Neural Network (CNN) that processes `256x256` tiles of the image and predicts text bounding boxes.70   * `recognizer_*.tflite` and `line_recognizer.fb` are CRNN (Convolutional Recurrent Neural Network) architectures that predict CTC logits for cropped text line images.712. **`*_label_map.pb` / `LabelMap.pb` (Vocabulary)**:72   * Binary Protobuf files mapping character indices to Unicode symbols for decoding CTC outputs.733. **`*_lm.compact_fst.gz` & `*.syms` (Language Models)**:74   * Compact Finite State Transducer (FST) language models and symbol mapping files. These are used for advanced Beam Search decoding, correcting spelling and character sequences based on word frequencies.754. **`*_prior.pb` / `*_config.pb` (Priors & Config)**:76   * Character prior probabilities and model configurations used to calibrate neural network outputs before applying the language model.77 78---79 80## 🚀 How to Use with `openmlkitOCR`81 82The Python package `openmlkitOCR` handles automatic downloading and caching of these models from Hugging Face if they are not present locally.83 84### 1. Installation85Install the library directly using pip:86```bash87pip install openmlkitOCR88```89 90### 2. Python Usage Example91```python92import os93import cv294from openmlkit import OpenMLKitOCR95 96# Configure the pipeline to pull models from this Hugging Face repository97os.environ["OPENMLKIT_MODEL_REPO"] = "0cve0/OpenMLKitOCR"98 99# Initialize the pipeline for a specific language (e.g., 'en' for English/Latin)100# This will automatically download and cache the detector and recognizer files.101ocr = OpenMLKitOCR(lang='en')102 103# Load an image104image = cv2.imread("test_image.jpg")105 106# Run OCR (detection & recognition)107results = ocr.run(image, score_threshold=0.35)108 109# Output the localized text bounding boxes and recognised characters110for item in results:111    print(f"Box: {item['box']} -> Text: {item['text']}")112```113 114---115 116## ⚖️ License and Disclaimer117 118* **Software**: The Python library [OpenMLkitOCR](https://github.com/0cve0/OpenMLkitOCR) is licensed under the **Apache 2.0 License**.119* **Model Weights**: The model weights and configurations in this repository are extracted from Google ML Kit APK components and are subject to Google's terms of service and license agreements. These models are intended for educational, research, and non-commercial local testing purposes.120