Article: Deep multiple affinity model for proposal-free single instance segmentation Journal: International Journal of Computational Vision and Robotics (IJCVR) 2024 Vol.14 No.5 pp.491 - 509 Abstract: We improve on an existing instance segmentation model with a probabilistic extension to the encoded neighbourhood branch model (Bailoni et al., 2020) - we call it multiple outputs encoded neighbourhood branch (mENB) model. The mENB predicts - for each voxel in a 3D volume, a distribution of central masks, where each mask represents affinities of its central voxel and the neighbouring voxels within the mask. When post-processed using a graph partition algorithm, these masks collectively delineates the boundaries of each instance of the target class within the input volume. Our algorithm is efficient - due to active learning, more accurate and it is robust to Gaussian noise and model weights perturbations. We conducted two experiments: 1) the first experiment compared mask predictions of our technique against the baseline (Bailoni et al., 2020) using the CREMI 2016 neuron segmentation dataset and the results showed a more accurate masks predictions with uncertainty quantification; 2) in the second experiment, we tested segmented instances against the popular proposal-based mask-RCNN and the results showed that our technique yields better precision and intersection over union. Inderscience Publishers - linking academia, business and industry through research

Title: Deep multiple affinity model for proposal-free single instance segmentation

Authors: Isah Charles Saidu; Lehel Csató

Addresses: Department of Computer Science, African University of Science and Technology, Nigeria; Faculty of Computing and Applied Science, Department of Computer Science, Baze University, Nigeria ' Faculty of Mathematics and Informatics, Babeș-Bolyai University of Cluj-Napoca, Romania; Faculty of Informatics, Eötvös Loránd University, Budapest, Hungary

Abstract: We improve on an existing instance segmentation model with a probabilistic extension to the encoded neighbourhood branch model (Bailoni et al., 2020) - we call it multiple outputs encoded neighbourhood branch (mENB) model. The mENB predicts - for each voxel in a 3D volume, a distribution of central masks, where each mask represents affinities of its central voxel and the neighbouring voxels within the mask. When post-processed using a graph partition algorithm, these masks collectively delineates the boundaries of each instance of the target class within the input volume. Our algorithm is efficient - due to active learning, more accurate and it is robust to Gaussian noise and model weights perturbations. We conducted two experiments: 1) the first experiment compared mask predictions of our technique against the baseline (Bailoni et al., 2020) using the CREMI 2016 neuron segmentation dataset and the results showed a more accurate masks predictions with uncertainty quantification; 2) in the second experiment, we tested segmented instances against the popular proposal-based mask-RCNN and the results showed that our technique yields better precision and intersection over union.

Keywords: segmentation; active learning; affinity model; uncertainty quantification.

DOI: 10.1504/IJCVR.2024.140817

International Journal of Computational Vision and Robotics, 2024 Vol.14 No.5, pp.491 - 509

Received: 04 Jun 2022
Accepted: 25 Oct 2022
Published online: 03 Sep 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Deep multiple affinity model for proposal-free single instance segmentation

Keep up-to-date