Segmentation by Region Growing

Image segmentation algorithms aim to break the input image into regions that correspond to objects or materials of interest.

In shoot phenotyping in the field, for example, a common first step is to separate plant material from soil. This may be done using any image features, though colour and texture (local image statistics) are the most common, and typically defines regions as physically connected sets of pixels that are in some sense similar.

While the pixels depicting real objects in natural images are likely to share properties, they are rarely perfectly uniform in terms of colour, etc. In the plant/soil segmentation problem the soil will broadly speaking be brown and the plants green, but each material will show a degree of variation and in some cases – perhaps as small weeds germinate – may appear mixed.

Region growing addresses this problem by first identifying small “seed” regions with a high degree of similarity and then expanding these across the image, only adding pixels which both lie on the boundary of an existing region and are considered similar enough to some summary properties of that region. The criteria used to define a seed and to grow a seed are typically different: seed regions might be required to have e.g. a very small standard deviation in grey level or hue, giving high confidence that this it arises from a single object. This restriction can, however, be relaxed when extending the region. If we are confident that a seed is, e.g. soil, it is more likely that neighbouring pixels with similar, if a little more varied colours are also soil. The seed provides context and confidence to the region growing step.

Region growing provides a framework for image segmentation, rather than a specific algorithm. The designer is free to define seed properties and growth criteria that suit the problem at hand, indeed there is no generic region growing tool. Region growing is particularly well-suited to interactive image segmentation. Here seeds are typically identified by a human user and grown by the algorithm, allowing the user to guide processing quickly and easily, without having to manually identify each pixel in a target object.