Converting Opensourced OCR Models to TFLite

This repository is to create tflite models for the available ocr models.

TFLite Models Available

• Captcha OCR
• Keras OCR
• Deep Text Recognition (EasyOCR uses the fine-tuned version of this model)

Note: Currently Keras OCR doesn't support Integer Quantization as of now and already reported to TFLite team.

About the files

├── colabs
    ├── KERAS_OCR_TFLITE.ipynb       --> Code for Convert Keras OCR model to TFLite and doing inference.
    ├── captcha_ocr_tflite.ipynb     --> Code to Convert Keras Captcha OCR to TFLite and code to do inference.
    ├── captcha_ocr_inference.ipynb  --> Inference code for captcha OCR.
    ├── Deep_Text_Recognition_TFLite.ipynb  --> Code to Convert Deep Text Recognition Model to TFLite with inference.
    ├── ENDTOEND_OCR_TFLITE.ipynb    --> End to End OCR with CRAFT,EAST as text detector and KERAS OCR as Recognition Engine.

├── models
    ├──  ocr_dr.tflite                --> Captcha OCR TFLite model.
    ├──  keras_ocr_dr.tflite          --> Keras OCR Dynamic Range Quantized TFLite Model
    ├──  keras_ocr_float16.tflite     --> Keras OCR Float16 Quantized TFLite Model
    ├──  cnn_dr.tflite                --> Deep Text Recognition Dynamic Range Quantized TFLite Model
    ├──  cnn_float16.tflite           --> Deep Text Recognition Float16 Quantized FLite Model 
├── images
    Contains sample input images used for inference.
├── data
    Contains representative dataset used while converting to TFLite using Integer Quantization.

Benchmarks

KERAS OCR

                                       All the above benchmarks are w.r.t Redmi K20 Pro using 4threads.

Tensorflow Hub

Keras OCR Models are published in TensorFlow Hub

You can find this repo for converting CRAFT models to TFLite and using them for inference. For more details please find these blogs on Text Detectors.

• Converting CRAFT to TFLite: A Guide to PyTorch-TFLite Conversion
• A Battle of Text Detectors for Mobile Deployments: CRAFT vs. EAST

Feel free to suggest any other models through issues or contribute through Pull Requests.