Tensorflow implementation of a Convolutive Autoencoder-like Nonnegative Matrix Factorization, with a Spatial Attention block.
The method has been developed to unmix MultiSpectral Images of ancient documents. The considered scenarios are :
- written text extraction
- inks distinction
- image segmentation, i.e. spectral unmixing, in a blind source separation context
The architecture of the proposed method is described by the following figures :
![]() |
![]() |
---|---|
Figure 1 : Structure of the proposed network | Figure 2 : Structure of the Spatial Attention block |
Preliminary results can be viewed in the following table :
Dataset | Image name | Extracted abundance maps | Ground Truth |
---|---|---|---|
MSBin [1] | BT6 | ![]() |
![]() |
MSTEx [2] | z802 | ![]() |
![]() |
The datasets used are introduced in :
[1] F. Hollaus, S. Brenner, et R. Sablatnig, « CNN Based Binarization of MultiSpectral Document Images », in 2019 International Conference on Document Analysis and Recognition (ICDAR), Sydney, Australia: IEEE, sept. 2019, p. 533‑538. doi: 10.1109/ICDAR.2019.00091.
[2] M. Kalacska, « ICDAR 2015 MultiSpectral Text Extraction Contest (MS-TEx 2015) », https://www.academia.edu/26113049/ICDAR_2015_MultiSpectral_Text_Extraction_Contest_MS_TEx_2015_
The used libraries are listed in requirements.txt
.