Abstract
Abstract In this thesis, the ability of segmentation of Distance Regularized Level Set Evolution (DRLSE) is investigated and improved by changing the definition of edge indicator and foreseeing the weak edge information. The Distance Regularized Level Set Evolution is based on the gradient flow that minimizes an energy functional with a distance regularization term and an external energy term that drives the motion of the zero level set (contour) toward desired locations. The DRLSE has good efficiency and accuracy in locating contours of objects for conventional images by maintaining the regularity of level set function with the distance regularization term. The stable condition is well satisfied by this term. But for some images with multiple objects separated by short distances or having deep jagged edges, the DRLSE may cause the zero level set driven to contours far away from the desired locations. Moreover, the proximity or deep jagged edges of objects affect external energy term such that re-tuning of the parameter is needed to be able to give additional external force to drive the motion of the contour, which then might result in errors in the evolution and even destruction of level set function. We found out that such situations could be remedied by adding more directions (45 and 135 degrees) for computing edge indicator to the DRLSE whose original edge indicator has only two directions (0 and 90 degrees). The change in the definition of edge indicator can not only keep the stable level set function but also drive the motion of the contour more correctly. Like DRLSE, our solution is sensitive to noise corrupting the desired object boundary, therefore we apply some relevant adjustments including bilateral filtering to smooth noise out while preserving edge and foreknowing the weak edge information by canny edge detector. Several experiments have been conducted to demonstrate and compare the segmentation capability of the original DRLSE and the proposed improved edge indicator DRLSE. In addition to subjective evaluation, some objective measure of errors (MOEs) such as ME (misclassification error)、RFAE (relative foreground area error)、NU (region non-uniformity)、MHD (modified Hausdorff distances)、EMM (edge mismatch) and mean error (contour model especially) are adopted to evaluate the efficacy of the proposed approach. From the experimental results, the proposed approach did provide an easy but efficient improvement to the original DRLSE. Keywords: image segmentation, level set, weak edge, external energy, edge indicator, bilateral filter, canny edge detector, measure of error