Comparison Between Equivalent Architectures of Complex-valued and Real-valued Neural Networks - Application on Polarimetric SAR Image Segmentation - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Journal of Signal Processing Systems Année : 2022

Comparison Between Equivalent Architectures of Complex-valued and Real-valued Neural Networks - Application on Polarimetric SAR Image Segmentation

Résumé

We present an in-depth statistical comparison among several Complex-Valued Neural Network (CVNN) models on the Oberpfaffenhofen Polarimetric Synthetic Aperture Radar (PolSAR) database and compare them against Real-Valued Neural Network (RVNN) architectures. The necessity to define the equivalence between the models emerges in order to compare both networks fairly. A novel definition for an equivalent-RVNN in terms of real-valued trainable parameters that maintain the aspect ratio is extended for convolutional layers based on previous work Barrachina et al. (2021 IEEE 31st International Workshop on Machine Learning for Signal Processing (MLSP), 2021). We illustrate that CVNN obtains better statistical performance for classification on the PolSAR image across a range of architectures than a capacity equivalent-RVNN, indicating that this behavior is likely independent of the model itself.
Fichier principal
Vignette du fichier
DEMR22017.pdf (1.96 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03771786 , version 1 (07-09-2022)

Identifiants

Citer

José Agustín Barrachina, Chengfang Ren, Christèle Morisseau, Gilles Vieillard, Jean-Philippe Ovarlez. Comparison Between Equivalent Architectures of Complex-valued and Real-valued Neural Networks - Application on Polarimetric SAR Image Segmentation. Journal of Signal Processing Systems, 2022, ⟨10.1007/s11265-022-01793-0⟩. ⟨hal-03771786⟩
32 Consultations
42 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More