Model Card for sbb_binarization
This is a pixelwise segmentation model for document image binarization. The model is a hybrid CNN-Transformer encoder-decoder model (Resnet50-Unet) developed by the Berlin State Library (SBB) in the QURATOR project. It can be used to convert all pixels in a color or grayscale document image to only black or white pixels. The main aim is to improve the contrast between foreground (text) and background (paper) for purposes of Optical Character Recognition (OCR).
Table of Contents
- Model Card for sbb_binarization
- Table of Contents
- Model Details
- Uses
- Bias, Risks, and Limitations
- Training Details
- Evaluation
- Model Examination
- Environmental Impact
- Technical Specifications
- Citation
- Glossary [optional]
- More Information [optional]
- Model Card Authors
- Model Card Contact
- How to Get Started with the Model
Model Details
Model Description
Document image binarization is one of the main pre-processing steps for text recognition in document image analysis. Noise, faint characters, bad scanning conditions, uneven light exposure or paper aging can cause artifacts that negatively impact text recognition algorithms. The task of binarization is to segment the foreground (text) from these degradations in order to improve Optical Character Recognition (OCR) results. Convolutional neural networks (CNNs) are one popular method for binarization, while Vision Transformers are gaining performance. The sbb_binarization model therefore applies a hybrid CNN-Transformer encoder-decoder model architecture.
- Developed by: Vahid Rezanezhad
- Shared by [Optional]: Staatsbibliothek zu Berlin / Berlin State Library
- Model type: Neural Network
- Language(s) (NLP): Irrelevant; works on all languages
- License: apache-2.0
- Parent Model: ResNet-50, see the paper by Zhang et al
- Resources for more information: More information needed
- GitHub Repo
- Associated Paper 1 Time-Quality Binarization Competition
- Associated Paper 2 Time-Quality Document Image Binarization
Uses
Document image binarization is the main use case of this model. The architecture of this model alongside with training techniques like model weights ensembling can reach or outperform state-of-the-art results on standard Document Binarization Competition (DIBCO) datasets in the both machine-printed and handwritten documents.
Direct Use
The intended use is the binarization of document images, particularly of historical documents, understood as one of the main pre-processing steps for text recognition.
Downstream Use
A possible downstream use of this model might lie with the binarization of illustrative elements contained in document images such as digitized newspapers, magazines or books. In such cases, binarization might support analysis of creator attribution, artistic style (e.g., in line drawings), or analysis of image similarity. Furthermore, the model can be used or trained for any other image enhancement use cases too.
Out-of-Scope Use
This model does NOT perform any Optical Character Recognition (OCR), it is an image-to-image model only.
Bias, Risks, and Limitations
The aim of the development of this model was to improve document image binarization as a necessary pre-processing step. Since the content of the document images is not touched, ethical challenges cannot be identified. The endeavor of developing the model was not undertaken for profit; though a product based on this model might be developed in the future, it will always remain openly accessible without any commercial interest. This algorithm performs a pixelwise segmentation which is done in patches. Therefore, one technical limitation of this model is that it is unable to capture and see long range dependencies.
Recommendations
The application of machine learning models to convert a document image into a binary output is a process which can still be improved. We have used many pseudo-labeled images to train our model, so any improvement or ground truth extension would probably lead to better results.
Training Details
Training Data
The dataset used for training is a combination of training sets from previous DIBCO binarization competitions alongside with the Palm Leaf dataset and the Persian Heritage Image Binarization Competition PHIBC dataset, with additional pseudo-labeled images from the Berlin State Library (SBB; datasets to be published). Furthermore, a dataset for very dark or very bright images has been produced for training.
Training Procedure
We have used a batch size of 8 with learning rate of 1e − 4 for 20 epochs. A soft dice is applied as loss function. In the training we have taken advantage of dataset augmentation. The augmentation includes flipping, scaling and blurring. The best model weights are chosen based on some problematic documents from the SBB dataset. The final model results of the ensemble of the best weights.
Preprocessing
In order to use this model for binarization no preprocessing is needed for the input image.
Speeds, Sizes, Times
More information needed
Training hyperparameters
In the training process, the hyperparameters were patch size, learning rate, number of epochs and depth of encoder part.
Training results
See the two papers listed below in the evaluation section.
Evaluation
In the DocEng’2021 Time-Quality Binarization Competition, the model ranked twelve times under the top 8 of 63 methods, winning 2 tasks.
In the ICDAR 2021 Competition on Time-Quality Document Image Binarization, the model ranked two times under the top 20 of 61 methods, winning 1 task.
Testing Data, Factors & Metrics
Testing Data
The testing data are the ones used in the Time-Quality Binarization Competition and listed in the paper on Time-Quality Document Image Binarization.
Factors
More information needed.
Metrics
The model has been evaluated both based on OCR and pixelwise segmentation results. The metrics which have been used in the case of visual evaluation are pixel proportion error and Cohen's Kappa value, and Levenshtein distance error in the case of OCR.
Results
See the two papers listed above in the evaluation section.
Model Examination
More information needed.
Environmental Impact
Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
- Hardware Type: Nvidia 2080.
- Hours used: Two days.
- Cloud Provider: No cloud.
- Compute Region: Germany.
- Carbon Emitted: More information needed.
Technical Specifications
Model Architecture and Objective
The proposed model is a hybrid CNN-Transformer encoder-decoder model. The encoder part consists of a ResNet-50 model. The ResNet-50 includes convolutional neural networks and is responsible for extracting as many features as possible from the input image. After that the input image goes through the CNN part, then the output undergoes upsampling convolutional layers until the same output size as in the input image is reached.
Compute Infrastructure
Training has been performed on a single Nvidia 2080 GPU.
Hardware
See above.
Software
See the code published on GitHub.
Citation
Coming soon.
BibTeX:
More information needed.
APA:
More information needed.
Glossary [optional]
More information needed
More Information [optional]
More information needed.
Model Card Authors
Vahid Rezanezhad, Clemens Neudecker, Konstantin Baierer and Jörg Lehmann
Model Card Contact
Questions and comments about the model can be directed to Clemens Neudecker at clemens.neudecker@sbb.spk-berlin.de, questions and comments about the model card can be directed to Jörg Lehmann at joerg.lehmann@sbb.spk-berlin.de
How to Get Started with the Model
Use the code below to get started with the model.
sbb_binarize
-m <from_pretrained_keras("sbb_binarization")>
- Downloads last month
- 0