Deep Convolutional Level Set Method for Image Segmentation

Abstract

Level Set Method is a popular method for image segmentation. One of the problems in Level Set Method is finding the right initial surface parameter, which implicitly affects the curve evolution and ultimately the segmentation result. By setting the initial curve too far away from the target object, Level Set Method could potentially miss the target altogether, whereas by setting the initial curve as general as possible – i.e. capturing the whole image – makes Level Set Method susceptible to noise. Recently, deep-learning methods, especially Convolutional Neural Network (CNN), have been proven to achieve state-of-the-art performance in many computer vision tasks such as image classification and detection. In this paper, a new method is proposed, called Deep Convolutional Level Set Method (DCLSM). The idea is to use the CNN object detector as a prior for Level Set Method segmentation. Using DCLSM it is possible to significantly improve the segmentation accuracy and precision of the classic Level Set Method. It was also found that the prior used in the proposed method is the lower and upper bound for DCLSM’s precision and recall, respectively.