Multi-scale graph representation learning for remote sensing image analysis

When:
30/11/2020 – 01/12/2020 all-day
2020-11-30T01:00:00+01:00
2020-12-01T01:00:00+01:00

Offre en lien avec l’Action/le Réseau : MACLEAN/– — –

Laboratoire/Entreprise : IRISA (OBELIX team)
Durée : 6 months
Contact : minh-tan.pham@irisa.fr
Date limite de publication : 2020-11-30

Contexte :
In the past few years, the amount of earth observation missions using remote sensing technologies has increased dramatically, providing a huge number of multimodal data coming from different sensors: optical, radar, lidar, etc. The need of efficient and reliable methods for multimodal remote sensing data analysis becomes crucial to exploit their complementary information for tackling various applications such as land-cover mapping and updating, scene understanding, urbanization trend detection and prediction, etc. Among modern techniques, object-based approach using graph model appears to be a promising solution.

Sujet :
This internship proposal aims at studying the ability of graph structures to model and characterize the spatial relationships of objects and regions from an image at different scales. That means we are interested to work on object/region levels, not the pixel level from the image. From the achieved graph structures, recent frameworks based on graph representation learning (e.g. graph convolutional neural networks, graph autoencoder) and graph distance metric learning could be investigated to perform structured graph embedding into robust feature spaces. For remote sensing applications, we are interested in various tasks including image retrieval, classification and scene matching, with applications to ecological or humanitarian challenges. More specifically, this work will concentrate on extracting meaningful spatial graphs that can be used for reasoning (akin to [5]). The work will leverage on theoretical works developed in the team [6] to design sensible loss between graphs objects to learn efficiently neural networks that will predict the graph structure.

Potential outcomes of the internship will lead to publications in remote sensing, computer vision or machine learning fields, depending on the nature of the contributions. Let us finally note that this internship will be part of the AI chair OTTOPIA funded by ANR (starting beginning of 2021), for which potential fundings are available for the candidate to enter a PhD track after the internship.

For more information: https://www-obelix.irisa.fr/files/2020/10/stageM2_graph_2020.pdf

Profil du candidat :
Student Master 2, Ecole d’Ingénieur or equivalent with excellent academic track;

Formation et compétences requises :
• Background in computer science and/or machine/statistical learning and/or applied mathematics for signal and image processing;
• Excellent programming in Python (familiar with one of deep learning packages, such as PyTorch or Tensorflow, is a must.)

Adresse d’emploi :
Université Bretagne Sud – IRISA (OBELIX team: https://www-obelix.irisa.fr/), Vannes 56000, France

Document attaché : 202010141228_stageM2_graph_2020.pdf