Increasing the parameter robustness of active contours using image data driven initializations

Although the well-known task of image segmentation which partitions the image into separated areas including different objects is part of almost every image processing application it still remains challenging. In the early 90's level set methods became a popular framework for front propagation methods like active contours (ACs) including edge-based and region-based models. Due to the optimization in a local manner those methods lead to segmentation results which depend on the initialization. While edge-based models are commonly known to be very sensitive to the initialization in noisy and realistic images, the initializing of region-based models are expected to be much more robust to varying initialization. In this paper we investigate the parameter robustness of different edge-based models concerning different initializations for synthetic and real images containing Gaussian noise with different noise levels. We show that the robustness of region-based ACs can be significantly increased by image data driven initializations. We compare the segmentation results of different models on synthetic and real images with respect to the Dice coefficient.