Deep Learning based Optimization of Extended Topological Active Net for Multi Object Segmentation

Abstract— Segmenting objects in image is an important task in many computer vision based applications. Activenets are a popular segmentation method which partitions the objects by creating holes in mesh to fit the object within the mesh. Different types of Activenets have been proposed in literature based on the optimization methods used for fitting the objects within mesh. Topological ActiveNets (TAN) and its extension Extended Topological ActiveNets (ETAN) are two popular Activenets applying energy based optimization. The problem in use of ETAN for the case of complex images is that it often leads to local optima. In this work, a deep learning based optimization is proposed to improve the ETAN and prevent it from the local optima problem. By this way, segmentation accuracy is improved. Keywords: TAN, ETAN, Deep learning and Objectiveness.