[Imageworld] One year post-doc position: Comparison of segmentation algorithms and texture analysis for leaf images adapted to smartphones

[Antoine Vacavant]

Hello,

please consider the following pot-doc position, that should start by
September or October 2012.

Sincerely,
Antoine Vacavant



Title:
Comparison of segmentation algorithms and texture analysis for leaf
images adapted to smartphones

Main topics:
Image segmentation; embedded image processing; evaluation of
segmentation algorithms; texture analysis; classification; multi-scale
algorithms

Place:
Le Puy-en-Velay (43), France

Duration & salary:
12 months, starting at September or October 2012, paid approximately 1500€/month

Associated project:
ANR Project ReVeS (Reconnaissance de Végétaux pour des interfaces
Smartphones), http://liris.cnrs.fr/reves/

Supervisor:
Antoine Vacavant, antoine.vacavant at iut.u-clermont1.fr

Context & motivation:
The segmentation of natural images, such as plants in their
environment, is a very complex task and a challenge in many image
analysis applications. In the ReVeS project [1], we are interested in
the analysis of photos of tree leaves taken with a smartphone. The
main purpose of ReVeS is to propose an embedded application able to
recognize the tree associated with the processed leaf.
A first pipeline has been proposed in [2,3] in order to carry out this
process, keeping in mind that it should be embedded inside mobile
phones. Our approach tackles the problem of  complex and uncontrolled
image background by using a model-based segmentation. We have shown
that classic segmentation algorithms like active contours, employed
without this a priori information, may lead to very noisy results, due
to the similarity between the object of interest and the background.
To ensure that our method leads to a high quality and competitive
result w.r.t. the literature, the main goal of this post-doc is to
study the comparison of segmentation algorithms, adapted to leaf
images and mobile phones.

Detailed goals:
First, the post-doc has to study the state of the art of segmentation
algorithms, and mainly (but not restricted to):
-multi-scale techniques (split & merge, multi-resolution, hierarchical, etc.);
-active contour or region growing methods (level sets, fast marching,
narrow band, etc.);
-clustering algorithms (k-means, etc.);
-model-based approaches.
Simultaneously, a survey of evaluation criteria should be conducted,
and should also include:
covering area, by considering a binary mask as ground truth [3];
-classification measures, as true/false positive/negatives,
F-measures, etc. as frequently used in background/foreground
segmentation algorithms in computer vision [4];
-contour based criteria, by using the polygonal contour of the ground
truth, such as Haussdorff distance,  error of tangents orientations,
etc. [5].
The post-doc should implement many selected algorithms on a smartphone
or a tablet, to ensure the possibility to use them in a mobile
application.
During this post-doc, the introduction of texture analysis -possibly
at several levels in a multi-scale scheme- should be studied to
improve the segmentation.
The use of texture could also be interesting for classification
issues, and could help to build a robust descriptor.

Required skills:
Besides knowledge on image/signal processing and algorithmic,
mathematics topics, the PhD should have good practice of C/C++ and or
Java programming (Java preferred). He/She would work with the DGtal
[6] and OpenCV libraries, and should have some basics on Android SDK
or an other SDK  for mobile applications.

References:
[1] ReVeS project (Reconnaissance de Végétaux pour des interfaces
Smartphones), http://liris.cnrs.fr/reves.
[2] Guillaume Cerutti, Laure Tougne, Antoine Vacavant and Didier
Coquin. A Parametric Active Polygon for Leaf Segmentation and Shape
Estimation. In ISVC 2011, LNCS 6938, 202-213,  2011.
[3] Guillaume Cerutti, Laure Tougne, Julien Mille, Antoine Vacavant
and Didier Coquin. Guiding Active Contours for Tree Leaf Segmentation
and Identification. In CLEF 2011.
[4] Yoann Dhome, Nicolas Tronson, Antoine Vacavant, Thierry Chateau,
Christophe Gabard, Yann Goyat, and Dominique Gruyer. A Benchmark for
Background Subtraction Algorithms in Monocular Vision: a Comparative
Study. In IPTA 2010.
[5] Antoine Vacavant, Tristan Roussillon and Bertrand Kerautret.
Unsupervised Polygonal Reconstruction of Noisy Contours by a Discrete
Irregular Approach. In IWCIA 2011, LNCS 6636, pp. 389-409, 2011.
[6] DGtal (Discrete GeometryTools and Algorithms), http://liris.cnrs.fr/dgtal/.

-- 
Antoine Vacavant
Maître de Conférences en Informatique
Responsable de la Licence Pro SIL Imagerie
--
IUT
8, Rue Jean-Baptiste Fabre
BP 70 219
43006 Le Puy en Velay - FRANCE
tél secrétariat : 04 71 09 90 88