Incorporating Flexibility into the Normalized Cut Image Segmentation Algorithm

<p>Image segmentation is the process of dividing a digital image into individual segments which share similar visual characteristics. The Normalized Cut (NCut) algorithm is one of the commonly used graph-based approaches in image segmentation. The NCut algorithm aims to extract big picture segments or global features of an image, a process which closely resembles how a human would approach image segmentation [1]. However, the algorithm is heavily dependent on a constant tuning parameter that is subject to arbitrary assignment prior to running the algorithm. This tuning parameter is independent of the image and indirectly specifies the level of details in the image one requires from the segmentation. Given this shortcoming, we propose a more flexible approach that introduces a local tuning parameter for each pixel over a small neighborhood in the image. We believe that the tuning parameter should represent the local variation of the features in the image in order to correctly tune the necessary components in the segmentation process. In particular, we look at improving the segmentations by introducing multiple "local-variation" tuning parameters that are adjusted to specific regions of the image. We do this through a semi-supervised method, where the regions are defined using the segmentations of the original NCut algorithm. Through our methodology, we incorporate additional local variation into tuning the algorithm without sacrificing the global features extracted by the original NCut algorithm. Results show that our methodology manages to improve the original NCut segmentations for some sample images.