Low-level Image Segmentation for a Vine Imaging Robot
Degree GrantorUniversity of Canterbury
Image segmentation is an important preprocessing step in most computer vision based applications, as it can significantly reduce future computation in tasks such as object classification. By grouping pixels that are similar with regard to a measure such as colour or position, classification can be performed on a per-segment basis, rather than per-pixel. This research examines several segmentation techniques and evaluates their performance at segmenting the network structure of vine images. Methods described in the literature are selected for comparison based on their performance at segmenting similar structures. The methods examined are k-means clustering, mean-shift clustering, normalised cuts segmentation, quadtree segmentation and watershed segmentation. We evaluate each method against five distinct images, based on their accuracy and efficiency at separating scene components such as vines, posts, wires and background. Evaluation is performed using a boundary-based comparison method to compare segmented images against hand generated ground truths. The clustering methods k-means and mean-shift are found to have the best performance. We propose mean-shift as the most suitable algorithm, due to its ability to produce a dynamic number of segments. We provide reasoning behind the relative successes and shortcomings of each method.