Introducing the Boundary-Aware loss for deep image segmentation - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Introducing the Boundary-Aware loss for deep image segmentation

Résumé

Most contemporary supervised image segmentation methods do not preserve the initial topology of the given input (like the closeness of the contours). One can generally remark that edge points have been inserted or removed when the binary prediction and the ground truth are compared. This can be critical when accurate localization of multiple interconnected objects is required. In this paper, we present a new loss function, called, Boundary-Aware loss (BALoss), based on the Minimum Barrier Distance (MBD) cut algorithm. It is able to locate what we call the leakage pixels and to encode the boundary information coming from the given ground truth. Thanks to this adapted loss, we are able to significantly refine the quality of the predicted boundaries during the learning procedure. Furthermore, our loss function is differentiable and can be applied to any kind of neural network used in image processing. We apply this loss function on the standard U-Net and DC U-Net on Electron Microscopy datasets. They are well-known to be challenging due to their high noise level and to the close or even connected objects covering the image space. Our segmentation performance, in terms of Variation of Information (VOI) and Adapted Rank Index (ARI), are very promising and lead to ∼15% better scores of VOI and ∼5% better scores of ARI than the state-of-the-art.
Fichier principal
Vignette du fichier
BMVC21_yizi.pdf (3.86 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03417244 , version 1 (05-11-2021)

Identifiants

  • HAL Id : hal-03417244 , version 1

Citer

Minh Ôn Vû Ngoc, Yizi Chen, Nicolas Boutry, Joseph Chazalon, Edwin Carlinet, et al.. Introducing the Boundary-Aware loss for deep image segmentation. British Machine Vision Conference (BMVC) 2021, Nov 2021, Virtual, United Kingdom. ⟨hal-03417244⟩
380 Consultations
397 Téléchargements

Partager

Gmail Facebook X LinkedIn More