“Lattice Cut” - Constructing Superpixels Using Layer Constraints

Image segmentation in computer vision refers to the process of dividing an image into multi-pixel, often irregular, contiguous regions. The resulting groups of pixels are then referred to as 'superpixels'  and can be used to give a different, more salient representation of an image.

One example use for image segmentation would be for facial recognition/tracking. Once the segments of an image can be divided into the two classes of face and non-face, other algorithms can be applied to determine whose face is present. Similarly, image segmentation is also used in medical imaging, for example, to measure the volume of an organ or to locate a tumor.

On a more theoretical note, one typical use for superpixel image segmentation is to find lines and boundaries in an image (as these will likely also be on the boundaries between the superpixel regions) or to assign a single, often more relevant, meaning to each region. In medical imaging, where a stack of images is frequently available, image segmentation leads to a similar stack of contours which can be used for 3D, volumetric reconstruction.

Segmentation is often an important first step in many image-processing algorithms and some practical applications for superpixel image segmentation are in object-location and object-class recognition in an image of a scene, face recognition and agricultural imaging (detecting diseased crops, for example).

Methods for image segmentation range from the simplest methods of thresholding to more 'intelligent,' learning-based methods. This paper provides a solution for globally-optimal image-segmentation which performs better than previous methods by using information about the boundaries of a superpixel as well as its contents.

Whilst the question of image segmentation has many good solutions, they are often application-dependent, particularly as segmentation forms only the first steps in an image processing problem. The lattice-based technique presented in this paper lends itself to further processing algorithms which also benefit from a regular image structure, with the advantage that the grid-like representation of the original pixels is somewhat preserved.