This thesis develops a new superpixel algorithm that displays excellent visual reconstruction of the original image. It achieves high stability across multiple random initialisations, achieved by producing superpixels directly corresponding to local image complexity. This is achieved by growing superpixels and dividing them on image variation. The existing analysis was not sufficient to take these properties into account so new measures of oversegmentation provide new insight into the optimum superpixel representation. As a consequence of the algorithm, it was discovered that CDS has properties that have eluded previous attempts, such as initialisation invariance and stability. The completely unsupervised nature of CDS makes them highly suitable for tasks such as application to a database containing images of unknown complexity. These new superpixel properties have allowed new applications for superpixel pre-processing to be produced. These are image segmentation; image compression; scene classification; and focus detection. In addition, a new method of objectively analysing regions of focus has been developed using Light-Field photography.