Update README.md
Browse files
README.md
CHANGED
@@ -8,10 +8,15 @@ tags:
|
|
8 |
- image-to-text
|
9 |
pipeline_tag: image-to-text
|
10 |
---
|
11 |
-
A CRNN model for Persian OCR. This model is based on a simple CNN + LSTM architecture inspired by [this paper](https://arxiv.org/abs/1507.05717).
|
12 |
-
our previous model [hezarai/crnn-base-fa-64x256](https://huggingface.co/hezarai/crnn-base-fa-64x256). The dataset for training this model was almost 5 times larger and the
|
13 |
-
maximum output length supported by this model has been increased from 32 to 48 characters. (The model can actually output 96 characters including blank but to tackle CTC decoding challenges no samples
|
14 |
-
longer than 48 characters have been fed to the model).
|
15 |
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
extract text boxes preferrably in word-level and then use this model), but it can be used to be fine-tuned on other domains like license plate or handwritten texts.
|
|
|
8 |
- image-to-text
|
9 |
pipeline_tag: image-to-text
|
10 |
---
|
11 |
+
A CRNN model for Persian OCR. This model is based on a simple CNN + LSTM architecture inspired by [this paper](https://arxiv.org/abs/1507.05717).
|
|
|
|
|
|
|
12 |
|
13 |
+
This is a successor model to our previous model [hezarai/crnn-base-fa-64x256](https://huggingface.co/hezarai/crnn-base-fa-64x256).
|
14 |
+
The improvements include:
|
15 |
+
- 5X larger dataset
|
16 |
+
- Change input image size from 64x256 to 32x384
|
17 |
+
- Increase max output length from 64 to 96 (Max length of the samples in the dataset was 48 to handle CTC loss issues)
|
18 |
+
- Support numbers and special characters (see id2label in `model_config.yaml`)
|
19 |
+
- Auto-handling of LTR characters like digits in between the text
|
20 |
+
|
21 |
+
Note that this model is only optimized for printed/scanned documents and works best on texts with a length of up to 50-ish characters. (For an end-to-end OCR pipeline, use a text detector model first to
|
22 |
extract text boxes preferrably in word-level and then use this model), but it can be used to be fine-tuned on other domains like license plate or handwritten texts.
|