Touching Digits

This database was first introduced by Oliveira et al in [1] (version 1.0). It was generated based on 2,000 isolated digits extracted from the hsf_0 series of NIST SD19. The main goal of this database was to provide a common catalog for evaluating segmentation algorithms. It is important to mention that the 2,000 images used to create it were correctly recognized by the our classifier (multi-layer perceptron that uses a 132-dimensional feature vector based on concavities and contour information). This issue is relevant for assessing the segmentation, and it will be further discussed in subsequent sections. The algorithm responsible for building the synthetic database is very simple, and is based on two rules:
  • It connects only digits produced by one writer. The information about the writer is provided in NIST SD19. Fifty different writers were considered.
  • The reference axis along which the digits slide is the center line.
The aim of these rules is to avoid unreasonable connections (e.g, very small digits connected to very big ones) and make the synthetic data more real.
According to the literature [2], touching digits can be classified into five different categories, as depicted in the figure below
This second version (V2.0), used in [3], of the database contains 79,466 samples distributed into the 100 classes of touching pairs, which correspond to the possible combinations of two digits. Some of the classes involving the digit 1 still contain fewer samples than other classes. Owing to the American style of handwriting, the digit 1 is very often with the other digit in the pair. The next table shows the distribution of the database based on the type of connection

Synthetic Data for the Segmentation-free Approach

In order to assess a segmentation-free approach based on Deep Convolutional Neural Networks [4], we have created some synthetic data containing touching strings of 2-, 3-, and 4-digits. The strings are built by concatenating isolated digits of NIST SD19 through the algorithm described in [3]. To avoid building a biased dataset, we have used the information about the authors available on the NIST SD19 so that digits from different authors were used exclusively for training, validation, and testing. The table below  shows the purpose (training, validation, and testing) and also the amount of data created.
Length/Classes Samples Authors from NIST Purpose
2-digit (100 classes) 161,563

53,907

55,091

 1000-1599

1600-1799

1800-1999

 Training

Validation

Testing
3-digit (1000 classes) 1,448,680

484,346

491,749

 1000-1599

1600-1799

1800-1999

 Training

Validation

Testing
4-digit * 100,000

20,000

20,000

 1000-1599

1600-1799

1800-1999

 Training

Validation

Testing

*The goal of 4-digit strings was to have data to train a classifier to predict the size of the numerical string (1,2,3, or 4 digits)

The architectures of the Convolutional Neural Network classifiers used in [4] are also available (based on the Caffe Framework)

How to obtain access to the images

Both Touching Digit (TP) database and Synthetic data for the segmentation-free approach may be used for non-commercial research provided you acknowledge the source of the image by citing the following papers in publications about your research:

Click here to download the datasets

References

[1] L. S. Oliveira, A. S. Britto Jr, and R. Sabourin. A synthetic database to assess segmentation algorithms. In 8th International Conference on Document Analysis and Recognition, pages 207– 211, 2005

[2] Y. K. Chen and J. F. Wang. Segmentation of single- or multiple-touching handwritten numeral string using background and foreground analysis. IEEE Trans. on Pattern Analysis and Machine Intelligence, 22(11):1304–1317, 2000.

[3] F. C. Ribas, L. S. Oliveira, A. S. Britto Jr, and R. Sabourin. Handwritten Digit Segmentation: A Comparative Study, International Journal of Document Analysis and Recognition, 16(2):127-137, 2013 .

[4] A. Hochuli, L. S. Oliveira, A. S. Britto Jr, and R. Sabourin. Handwritten Digits Segmentation: Is it still necessary? Pattern Recognition,  78:1-11, 2018.

[5] A. Hochuli, L. S. Oliveira, A. S. Britto Jr, and R. Sabourin. Segmentation-Free Approaches for Handwritten Numerical String Recognition, Int. Joint Conference on Neural Networks (IJCNN),  2018.

This database is licensed under a Creative Commons Attribution 4.0 International License.