Department of Computer Engineering
CS 590 SEMINAR
Topology-preserving and Shape-preserving Loss in Image Segmentation
Computer Engineering Department
Fully Convolutional Neural Networks (FCNNs) are used for image segmentation to classify each pixel from the given image. Widely used loss functions in FCNNs such as cross-entropy loss and dice loss ignore the topology and shape by only using the information from the respective pixel. Object-wise accuracies can be low while pixel-wise accuracies are high when using pixel-wise losses. There are proposed losses to overcome this problem: topology-preserving loss and shape-preserving loss. Since different shapes can have the same persistent homology in certain filtrations, topology-preserving loss and shape-preserving loss focus on two different information about segmentation maps. In this research, we will try to combine topology and shape signatures in one loss to exploit both topology and shape features.
DATE: 09 December 2019, Monday @ 15:40