Encodage de matrices de covariance par les vecteurs de Fisher log-euclidien : application à la classification supervisée d'images satellitaires

Abstract : This paper introduces a new hybrid architecture based on Fisher vector encoding (VF) of the convolutional layer outputs of a neural network. The originality of this work is based on the exploitation of second-order statistics via the calculation of local covariance matrices. Considering the intrinsic properties of the Riemannian manifold of covariance matrices, we propose to use the log-euclidean metric in order to extend the concept of VF encoding: the log-euclidean Fisher vectors (LE VF). The proposed architecture is then evaluated on different remote sensing databases : the UC Merced Land Use Land Cover database, the AID database, as well as on two Pléiades datasets on maritime pine forests and oyster beds.
Document type :
Conference papers
Complete list of metadatas

Cited literature [16 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-02294885
Contributor : Lionel Bombrun <>
Submitted on : Monday, September 23, 2019 - 5:52:07 PM
Last modification on : Saturday, September 28, 2019 - 9:50:24 PM

File

Akodad19_GRETSI.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02294885, version 1

Citation

Sara Akodad, Lionel Bombrun, Yannick Berthoumieu, Christian Germain. Encodage de matrices de covariance par les vecteurs de Fisher log-euclidien : application à la classification supervisée d'images satellitaires. Groupe d'Etudes du Traitement du Signal et des Images (GRETSI), Aug 2019, Lille, France. ⟨hal-02294885⟩

Share

Metrics

Record views

34

Files downloads

17