ICDAR2009 Handwriting Segmentation Contest

Click on thumbnails to see the actual images.

Method Description for Line Segmentation

First the input image is sub-sampled by a factor of four in both dimensions while turning it into a gray-level image:

Olena-thumbnail small.png

Then an anisotropic Gaussian filtering is applied (mainly horizontal). In the resulting fuzzy image, lines appear as dark shapes, that are regional minima:

Olena-thumbnail fuzzy.png

The morphological watershed transform is computed, leading into a partition of the image into regions. Those regions form an over-segmentation of text lines. In the picture below, the watershed line is depicted in red, superimposed over the fuzzy image.

Olena-thumbnail watershed.png

To obtain the segmentation into lines, a simple merging procedure is run on the region adjacency graph. Below, lines are colorized and depicted over the resized input image; the watershed line is shown in white.

Olena-thumbnail lines.png

Eventually a cleaning routine is applied to provide the final labeling image.

Method Description for Word Segmentation

First the input image is downsized and inverted.

Olena-thumbnail small.png

Second a morphological structural closing is applied with a rectangle as structural element.

Olena-thumbnail closing.png

From the input image a distance map to characters is computed. An attribute morphological closing is then applied to merge altogether non-significant regional minima.

Olena-thumbnail dmap.png

Last the morphological watershed transform is computed to obtain the segmentation into words. The watershed line is depicted below in red over the reduced input image.

Olena-thumbnail watershed.png

Source Code

The source code will be available on June 2009...