Beyond the Pixel-Wise Loss for Topology-Aware Delineation

摘要: Delineation of curvilinear structures is an important problem in Computer Vision with multiple practical applications. With the advent Deep Learning, many current approaches on automatic delineation have focused finding more powerful deep architectures, but continued using habitual pixel-wise losses such as binary cross-entropy. In this paper we claim that alone are unsuitable for because their inability to reflect topological impact mistakes final prediction. We propose a new loss term aware higher-order features linear structures. also exploit refinement pipeline iteratively applies same model over previous refine predictions at each step, while keeping number parameters and complexity constant. When combined standard loss, both our iterative boost quality predicted delineations, some cases almost doubling accuracy compared classifier trained cross-entropy alone. show approach outperforms state-of-the-art methods wide range data, from microscopy aerial images.