The evolution of snake toward automation for multiple blob-object segmentation

For the last two decades “active contour” or “snake” has been effective as an interactive image segmentation tool in a wide range of applications, especially for blob-object delineation. In the interactive snake segmentation process, a user draws a rough object outline; next, a cost function is optimized to drive the user-drawn contour a.k.a. snake to delineate the desired object boundary. Although successful as an interactive segmentation tool, snake exhibits poor performances in various noteworthy image segmentation applications that require complete automation. Examples include oil sand particle delineation, biological cell segmentation and so on.;This thesis presents a novel, completely automated snake/active contour algorithm for multiple blob-object delineation. The algorithm consists of three sequential steps: (a) snake initialization: where we apply a Probabilistic Quad Tree (PQT) based approximate segmentation technique on an image to find the regions of interest (ROI) where the probability of having objects is very high and place seeds uniformly within the ROIs; (b) snake evolution: where we evolve one novel interleave directional gradient vector flow (IDGVF) snake from each seed; (c) snake validation: where we classify the snakes into object and non-object classes using a novel adaptive regularized boosting (ARboost). Existing efforts towards snake automation have concentrated only on the succession of initialization and evolution steps and have practically overlooked the snake validation step. Here, we emphasize that we cannot skip the validation step, even though the initialization and evolution have performed well. Our proposed novel validation step, executed after complete convergence of a snake contour from a given initialization, classifies the evolved contour into desired object and nonobject classes.;ARboost employs a novel loss function for boosting that enables to classify snakes more accurately into object and non-object classes than other variants of boosting. PQT generates substantially fewer seed points and is therefore more efficient than other initialization methods without degrading the segmentation performance. We have demonstrated that IDGVF is more robust to initialization and it possesses a broader capture range than other variants of GVF snakes.;The proposed automated snake algorithm has been successfully applied to two real data sets: oil sand ore images that have relevance in the oil sand mining industry and leukocyte images that are significant in biomedical engineering.