DiTo97 / neural-binarization

A toolkit for efficient document image binarization (DIB) as a semantic segmentation problem

Binarization Segformer

A semantic segmentation model for pixel-wise document image binarization.

TODOs

• fine-tune Segformer on 1024 $\times$ 1024 images;
• set reduce_labels=True in Segformer processor to ignore the background;
• compare valid DIBCO metrics with SauvolaNet's paper.

Overview

Segformer is an efficient semantic segmentation model introduced by Xie et al. in 2021.

In this repository, we will provide a fine-tuning of Segformer for pixel-wise document image binarization.

Dataset

The dataset is an ensemble of 14 datasets replicating the setting used in SauvolaNet by Li et al. in 2021.

Figure 1. An example pair from the Bickley diary dataset

For more information on the dataset, see SauvolaNet's official repository.