Geometry and Variational Prior for Deep Learning based Image Segmentation: Dr Xiangyue Wang and Jun Liu (Beijing Normal University, China)
Wednesday 21st October 2020
09:00 via Zoom
Abstract: Convolutional Neural Networks (CNN) can well extract the features from natural images. However, the classification functions in the existing network architecture of CNNs are always simple and lack capabilities to handle important spatial information in a way that have been done for many well-known traditional variational models. Priors such as spatial regularization, volume, object shapes, topology priors cannot be well handled by existing CNN architectures. We propose a novel Soft Threshold Dynamics (STD) based framework which can easily integrate many priors such as local and nonlocal image edges information, star/convexity shapes of the classic variational models into the DCNNs for image segmentation. The novelty of our method is to interpret the activation functions (including softmax, sigmoid, ReLU) as primal-dual variational problem, and thus many priors can be imposed in the dual space. By unrolling method, we can build several STD based network architectures which can enable the outputs of CNN to have many special priors. The proposed method is a general framework and it can be applied to any image segmentation CNNs. We will give some applications to show the efficiency of our method.