fr | en
Laboratoire Angevin de Recherche en Ingénierie des Systèmes
  • Index
  • >News
  • >Defense of Ms. Hadmani GARBOUGE's thesis

Defense of Ms. Hadmani GARBOUGE's thesis9:00 am | IRHS | Meeting Room (rdc) | 42, rue Georges Morel 49070 BEAUCOUZE

Subject :  Deep learning applied to multi-component imagery for variety testing problems.

Director of thesis : Mr. David ROUSSEAU

Abstract

The thesis proposes original methodological contributions based on computer vision and machine learning techniques to the variety testing research. Variety testing consists in performing measurements to validate the quality and originality of any new variety before allowing its commercialization. The current tests are essentially the result of visual inspections, and digital phenotyping is not common. The work of this PhD has focused on developing automated variety testing methods for crops. We build new multicomponent imaging systems based on low-cost sensors and deep learning networks.
In the first part, we explore the potential of multi-component RGB-Depth sensors in variety testing, which provide trichromacy and distance informations from the plants to the camera. Early, intermediate, and late fusions on these components are examined using a convolutional, local memory and transformer neural network. We demonstrate the benefits of the distance map, especially for estimating the kinetics of the individual developmental stages of seedlings during the daytime and nighttime. Then, we explore the same approaches of RGB-Depth imaging for detecting developmental stages in small plots as textures.
In the second part, we address the issue of the possible transfer learning of traits measured in controlled environments (growth chamber, phytotron) to less controlled environments (greenhouses or fields). To do so, we revisit the detection of seedling development stages. Furthermore, a data augmentation method simulating shadows is proposed and shows interest in transfer learning approaches in greenhouses and the field. Lastly, using synthetic data for transfer learning is proposed.
In the third part, we develop an optimized multispectral camera for detecting and quantifying disease in plant resistance tests. Each step is detailed and validated on an image dataset acquired during three seasons. A global pipeline is presented, including deep learning and classical machine learning methods. An opening toward the automatic determination of acquisition protocol is proposed in Annex A.
Additionally to our methodological contributions, we provided various computer tools. We developed software to process RGBDepth image sequences for stage detection, measuring plant heights in real-time, etc. Furthermore, we provided annotated databases and generated trained models.

Scroll