Postdoctoral Fellowship Algorithms for The - Nancy, France - INRAE UMR Silva

Description

The Silva Joint Research Unit brings together people from AgroParisTech, INRAE, and the Université de Lorraine to conduct multidisciplinary research on wood, trees and forest ecosystems.

The main scientific goal of the UMR Silva is to develop pure and applied research to answer questions of society, including forest managers, about (1) the role and the future of forest ecosystems in the context of global changes, including climate change, and (2) the future of the timber industry, particularly in the Grand - Est region for which it is of major economic importance.

Site web:


Poste et missions:

This postdoctoral study addresses the log traceability problem and aims to develop new methods using image processing techniques and deep-learning approaches.

The size and shape of the cross-sections should be taken into account to facilitate matching (Charwat-Pessler et al., At the base of oak logs, the shape of the cross-section can be quite irregular.

Moreover, many angular points are visible, related to the logging of the tree, and could help to identify the logs However, the size and shape information will probably not be sufficient.

The pattern formed by the growth rings, more or less visible depending on the texture of the surface, could be exploited.

But unlike fingerprints, rings do not haveminutiae 2 (except the pith, at the centre of annual rings, which is a particular point) and only certain methods can be used, including those based on texture analysis (Schraml et al., 2015a).

One of the difficulties could be to extract this pattern, or parts of this pattern, from very rough, unprepared surfaces.

If the ring pattern is an important element for the identification of the logs, it will also be possible to use other characteristics such as the pith location (Decelle et al., 2021), the sawing marks, the width of the sapwood (Decelle et al., area of different colouration visible at the periphery of a cross-section under bark, lighter or darker depending on the case) and all other singularities that may be present on the surface.

One question will be which singularities can be used? For example, cracks may evolve over time and are not necessarily a good criterion.

Cutting marks could be used because they bring a specific identity to the log but this remains to be discussed.

Adapted methods of existing algorithms for fingerprint or iris identification have already been successfully tested at the University of Salzburg (Schraml et al., 2015b) on images taken from the log-ends of softwood logs.

However, the most recent work has shown that better performances can be obtained with convolutional neural networks (CNN) using a triplet loss function for training the network (Wimmer et al., 2021).

Mobilité géographique:

• Nationale
  Prise de fonction:
• 01/06/2023
  Profil:

PhD in Computer Science. Good programming skills (C++ and/or Python), experience with image analysis and processing, and scientific writing skills are expected. A good knowledge of CNN would be appreciated.


Objectifs:

The first step will be to understand and test on oak log-ends the algorithms developed by our Austrian colleagues for softwoods.

Very quickly it will be necessary to develop methods capable of identifying oak logs. Perhaps by allowing the use of a certain number of singularities at first and then restricting them later. Several types of algorithms could be considered, including adaptations of the most efficient CNN.

For this purpose, a database of 10,000 to 20,000 images of oak log-ends, which is currently being built up, can be exploited.