**2.2.3 Image segmentation**

Segmentation means separation of the image into interesting sub-regions for further analyses. For segmentation of 3D reconstructions, there is a large variety of algorithms. To mention few the most commonly used, are global algorithms like thresholding, spatially aware algorithms like region grow and watershed (Gonzales & Woods, 2002), optimisation based algorithms like level sets (Osher &, Sethian, 1988) or active contours (Kass et al, 1987), shape -based methods like LOG or DOG (Gonzales & Woods, 2002), texture -based methods like local binary patterns (Ojala et al, 1994), and combinations of these. Often the complexity of the required algorithm is directly proportional to the signal-to-noise ratio of the original image.

A typical task in segmentation procedure is to separate the image of a porous media to solid and void phase. In optimal conditions, the phases can be distinguished by the difference in their grey value distributions. However, the grey value distributions are often overlapping due to imaging noise. In many cases, thresholding combined with post processing gives satisfactory results. The post processing procedures aim to remove the non-connected parts, i.e. the "levitating" solid objects from the images or to fill the small non-effective (isolated) pores. In the cases where the overlapping regions of the grey value distributions are wide, more sophisticated algorithms are needed.

Fig. 2. An example of a noisy CXµT image (a) and the intensity profile plot for the area

In addition to ESF, the tomographic reconstructions are often contaminated by imaging noise. Collecting photons with Charge-Coupled Device (CCD) is a time-dependent discrete procedure that presents Poisson noise into the collected data. Utilisation of an analogue-todigital -converter such as in CCD causes Gaussian -type of noise into the images. In the final 3D representation of the tomographic sample, noise can be seen as random variation of grey values. This effect causes edge blurring. Furthermore, when binarised by thresholding

Many algorithms have been developed to decrease the noise, e.g. anisotropic diffusion (Perona & Malik, 1987), bilateral filtering (Tomasi & Manduchi, 1998) and SUSAN filtering (Smith & Brady, 1997). However, none of the filtering methods is perfect and post

Segmentation means separation of the image into interesting sub-regions for further analyses. For segmentation of 3D reconstructions, there is a large variety of algorithms. To mention few the most commonly used, are global algorithms like thresholding, spatially aware algorithms like region grow and watershed (Gonzales & Woods, 2002), optimisation based algorithms like level sets (Osher &, Sethian, 1988) or active contours (Kass et al, 1987), shape -based methods like LOG or DOG (Gonzales & Woods, 2002), texture -based methods like local binary patterns (Ojala et al, 1994), and combinations of these. Often the complexity of the required algorithm is directly proportional to the signal-to-noise ratio of the original

A typical task in segmentation procedure is to separate the image of a porous media to solid and void phase. In optimal conditions, the phases can be distinguished by the difference in their grey value distributions. However, the grey value distributions are often overlapping due to imaging noise. In many cases, thresholding combined with post processing gives satisfactory results. The post processing procedures aim to remove the non-connected parts, i.e. the "levitating" solid objects from the images or to fill the small non-effective (isolated) pores. In the cases where the overlapping regions of the grey value distributions are wide,

processing is often necessary to reduce the artefacts from the binarised images.

marked on the CXµT image (b).

**2.2.3 Image segmentation** 

image.

procedure, falsely labelled voxels can appear.

more sophisticated algorithms are needed.

We have developed a so-called forest fire algorithm to separate the material phases when their grey value distributions are overlapping too much for simpler methods. As an input, a user will give the limits for the overlapping area in the grey value histogram. The algorithm processes the grey values in between the given limits and decides whether it is solid or void by adding more voxels to each phase iteratively. In addition to grey value information, spatial information is incorporated. The voxels are added into the group if there are enough members of the same group around it. By adjusting the number of required neighbours, the sensitivity of the method can be adjusted individually for each phase. In practice, the method is closely related to region grow method, but it enhances the traditional region grow by adding a weak "surface tension" to it. The benefits of the method are smooth surfaces and possibility to alter the volume of selected phase by allowing either solid or void to conquer its area easier. The disadvantage of the method is that some of the smallest details can be lost. The name forest fire comes from simple forest fire simulations where forest is divided into cells which will catch fire if certain amounts of its neighbouring cells are already burning.
