v1.4

README.md

@@ -11,8 +11,14 @@ Ready-to-use OCR with 80+ [supported languages](https://www.jaided.ai/easyocr) a
 [Try Demo on our website](https://www.jaided.ai/easyocr)
 
 ## What's new
+- 29 June 2021 - Version 1.4
+    - [Instruction](https://github.com/JaidedAI/EasyOCR/blob/master/custom_model.md) on training/using custom recognition model
+    - Example [dataset](https://www.jaided.ai/easyocr/modelhub) for model training
+    - Batched image inference for GPU (thanks [@SamSamhuns](https://github.com/SamSamhuns), see [PR](https://github.com/JaidedAI/EasyOCR/pull/458))
+    - Vertical text support (thanks [@interactivetech](https://github.com/interactivetech)). This is for rotated text, not to be confused with vertical Chinese or Japanese text. (see [PR](https://github.com/JaidedAI/EasyOCR/pull/450))
+    - Output in dictionary format (thanks [@A2va](https://github.com/A2va), see [PR](https://github.com/JaidedAI/EasyOCR/pull/441))
 - 30 May 2021 - Version 1.3.2
-    - Faster greedy decoder (thanks [@samayala22](https://github.com/samayala22)) 
+    - Faster greedy decoder (thanks [@samayala22](https://github.com/samayala22))
     - Fix bug when text box's aspect ratio is disproportional (thanks [iQuartic](https://iquartic.com/) for bug report)
 - 20 April 2021 - Version 1.3.1
     - Add support for PIL image (thanks [@prays](https://github.com/prays))
@@ -28,8 +34,7 @@ Ready-to-use OCR with 80+ [supported languages](https://www.jaided.ai/easyocr) a
 - [Read all released notes](https://github.com/JaidedAI/EasyOCR/blob/master/releasenotes.md)
 
 ## What's coming next
-- Custom models
-- [New language support](https://github.com/JaidedAI/EasyOCR/issues/91)
+- Handwritten text support
 
 ## Examples
 
@@ -113,15 +118,14 @@ For more information, read [tutorial](https://www.jaided.ai/easyocr/tutorial) an
 $ easyocr -l ch_sim en -f chinese.jpg --detail=1 --gpu=True
 ```
 
+## Train/use your own model
+
+[read here](https://www.jaided.ai/custom_model.md)
+
 ## Implementation Roadmap
 
-1. Language packs: Expand support to more languages. We are aiming to cover > 80-90% of world's population. Also improve existing languages.
-2. Better documentation and api
-3. Language model for better decoding
-4. Handwritten support: The key is using GAN to generate realistic handwritten dataset.
-5. Faster processing time: model pruning (lite version) / quantization / export to other platforms (ONNX?)
-6. Open Dataset and model training pipeline
-7. Restructure code to support swappable detection and recognition algorithm.
+- Handwritten support
+- Restructure code to support swappable detection and recognition algorithm.
 The api should be as easy as
 ``` python
 reader = easyocr.Reader(['en'], detection='DB', recognition = 'Transformer')

custom_model.md

@@ -0,0 +1,30 @@
+# Custom model
+
+## How to train your custom model
+
+There are 2 options to train your own recognition model.
+
+**1. Open-source approach**
+
+For the open-source approach, you can use your own data or generate your own dataset. To generate your own data, we recommend using
+[TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator). We provide an example of dataset [here](https://jaided.ai/easyocr/modelhub/).
+After you have a dataset, you can train your own model by following this repository
+[deep-text-recognition-benchmark](https://github.com/clovaai/deep-text-recognition-benchmark).
+The network needs to be fully convolutional in order to predict flexible text length. Our current network is 'None-VGG-BiLSTM-CTC'.
+Once you got your trained model (.pth file), you need 2 additional files describing recognition network architecture and model configuration.
+The example is provided in `custom_example.zip` file [here](https://jaided.ai/easyocr/modelhub/).
+
+Please do not create an issue about data generation and model training in this repository. If you have any question regarding data generation and model training, please ask in the respective repositories.
+
+**2. Web-based approach**
+
+Jaided AI provides a web-based (paid) service for training your own model [here](https://jaided.ai/). You can train your model on the cloud and export it for local deployment. All 3 files are downloadable once model is trained on cloud.
+
+## How to use your custom model
+
+To use your own recognition model, you need 3 files either from open-source approach or web-based approach. These three files have to share the same name (for example, yourmodel.pth, yourmodel.yaml, yourmodel.py) and you will call your model by this name in EasyOCR api.
+
+We provide [custom_example.zip](https://jaided.ai/easyocr/modelhub/)
+as an example. Please download, extract and place `custom_example.py`, `custom_example.yaml` in user_network_directory (default = `~/.EasyOCR/user_network`) and place `custom_example.pth` in model directory (default = `~/.EasyOCR/model`)
+Once you place all 3 files in the right place. You can use `custom_example` by
+specifying `recog_network` like this `reader = easyocr.Reader(['en'], recog_network = 'custom_example')`.

easyocr/__init__.py

@@ -1,3 +1,3 @@
 from .easyocr import Reader
 
-__version__ = '1.3.2'
+__version__ = '1.4'

releasenotes.md

@@ -1,3 +1,9 @@
+- 29 June 2021 - Version 1.4
+    - [Instruction](https://github.com/JaidedAI/EasyOCR/blob/master/custom_model.md) on training/using custom recognition model
+    - [Example dataset](https://www.jaided.ai/easyocr/modelhub)
+    - Batched image inference for GPU (thanks [@SamSamhuns](https://github.com/SamSamhuns), see [PR](https://github.com/JaidedAI/EasyOCR/pull/458))
+    - Vertical text support (thanks [@interactivetech](https://github.com/interactivetech)). This is for rotated text, not to be confused with vertical Chinese or Japanese text. (see [PR](https://github.com/JaidedAI/EasyOCR/pull/450))
+    - Output in dictionary format (thanks [@A2va](https://github.com/A2va), see [PR](https://github.com/JaidedAI/EasyOCR/pull/441))
 - 30 May 2021 - Version 1.3.2
     - Faster greedy decoder (thanks [@samayala22](https://github.com/samayala22))
     - Fix bug when text box's aspect ratio is disproportional (thanks [iQuartic](https://iquartic.com/) for bug report)

setup.py

@@ -17,7 +17,7 @@ def readme():
     name='easyocr',
     packages=['easyocr'],
     include_package_data=True,
-    version='1.3.2',
+    version='1.4',
     install_requires=requirements,
     entry_points={"console_scripts": ["easyocr= easyocr.cli:main"]},
     license='Apache License 2.0',