Experiments with Skeletal Similarity for Document Binarization

This page presents the experiments made for investigating the use of Skeletal Similarity metrics (proposed by cite) to evaluate the performance of document binarization algorithms. These experiments were presented at the International Conference on Handwritten Recognition 2020 (ICFHR 2020), in the paper Skeletal Similarity based Structural Performance Evaluation for Document Binarization, by Silva et al.

In order to evaluate the use of these metrics on Document Binarization, we manually generated images that had small distortions which greatly affected pixelwise metrics but did not impair text readability. Below, we display an example image and the manually generated distortions:

Ground Truth	Dilated Image	Eroded Image
Slightly Translated Image	Noise Added Image	Missing Lines and Columns

Every manually generated distorted image from the experiment is available in the following link:
http://vision.ime.usp.br/~augustocms/ICFHR2020/manually_generated/
in which the files are named with the following pattern:

id_type_distortion.ext

id = number representing the image
distortion = manually generated distortion
type = img for image file, metrics for a file with the calculated metrics
ext = file extension

For further investigation of the usage of Skeletal Similarity metrics in document binarization, we also binarized 10 images from Document Image Binarization Competition 2018 (DIBCO) with real binarization algorithms (Otsu's, Sauvola's, Howe's and Zaragosa's). An example of an image binarized by this algorithms is shown below:

Original Image	Ground Truth Image	Otsu's binarization
Sauvola's binarization	Howe's binarization	Zaragosa's binarization

Every binarized image from the experiment is available in the following link:
http://vision.ime.usp.br/~augustocms/ICFHR2020/real_binarizations/
in which the files are named with the following pattern:

id_type_method.ext

id = number representing the image
method = binarization algorithm used
type = img for image file, metrics for a file with the calculated metrics
ext = file extension

If these experiments are useful to your research, please cite:
to be published