I do not consider this method suitable for a general purpose segmenter as the superpixel segmentation is not suitable for non-Latin scripts and semantic segmentation still requires the whole traditional line extraction pipeline.
A quick and dirty reimplementation of the shallow CNN superpixel labelling method for semantic segmentation presented in [0].
extract_patches.py takes a directory with jpg source images and png pixel
labelled images and extracts 28x28 image patches oriented around the centroids
of a SLIC oversegmentation. The class of each patch is derived from the highest
assigned class of the centroid pixel (alternatively there is a frequency
selection method implemented). Output is written to data/{0,1,2,3}/$uuid.png
main.py is used for training. Training is quite fast (~30s/100k samples) on a
modern machine.
Download the DIVA HisDB from [1] image files and pixel-labelled image files.
Pixel labellings/semantic segmentation (really a misnomer) are not immediately useful for line extraction for recognition, as the line separation/instance detection problem is not solved by these systems. The real world use of this approach requires a good binarization to project the predicted superpixel classes onto foreground pixels and an actual extractor such as a seam carver or the less-than-stellar one in kraken. The main reduction in complexity comes from being able to separate interlinear components, effectively increasing line spacing in challenging manuscripts.
- momentum in SGD.
[0] Chen, Kai, et al. "Convolutional neural networks for page segmentation of historical document images." Document Analysis and Recognition (ICDAR), 2017 14th IAPR International Conference on. Vol. 1. IEEE, 2017. [1] http://diuf.unifr.ch/main/hisdoc/diva-hisdb