Camille Kurtz
PhD Candidate in Computer Science
Original image
Segmented image
Object-based classified image
Camille KurtzPhD Candidate in Computer Science |
|
|
|
Master (M.Sc.) in Computer Science from the University of Strasbourg (2009)
Master manuscript (in French) : Master thesis 
(pdf 10.2 Mb)
Current position (since Oct. 2009) :
I am currently a PhD Candidate at the Image Sciences, Computer Sciences and Remote Sensing Laboratory (LSIIT), Strasbourg (France). My areas of expertise are image analysis and data mining in the remote sensing field. My PhD work is focused on multiresolution clustering, hierarchical image segmentation.
I am also a Teaching Assistant at the UFR de Mathématique et d'Informatique. I taught around 150h(+) to undergraduate computer science students (Distributed computing, Computer architecture, Graphs and operational computation, etc.).
Research interests :
Data mining
Clustering
Pattern recognition
Multiresolution approaches
Knowledge discovery
Image processing / analysis
Hierarchical segmentation
Mathematical morphology
Computer vision
Remote sensing
Referee :
Curriculum vitae :
(pdf 0.8 Mb)My Linked in profile
, my page on scopus
and my page on DBLP.
Phone : +33(0)3 68 85 45 78
Email : ckurtz_at_unistra_dot_fr
Postal address :
KURTZ Camille
LSIIT - UMR 7005
Pôle API
Bd Sébastien Brant
BP 10413
67412 Illkirch CEDEX FRANCE
Extraction of hierarchical patterns from multiresolution remote sensing images
In the field of remote sensing image analysis, the extraction of complex patterns from Very High Spatial Resolution optical images presents several challenges related to the size, the accuracy and the complexity of the considered data. Indeed, due to the large amount of ground details provided by these images, the classical approaches (i.e., the object-based ones) do not provide satisfactory results.
Based on the availability of several optical images of a same scene at various resolutions (Medium to Very High Spatial Resolution), we propose in this work a hierarchical framework to progressively extract segments/objects of interest from the lowest to the highest resolution data, and then finally determine complex patterns from VHSR images. This approach, inspired by the principles of photo-interpretation and human vision, has for purpose (1) to exploit the homogeneous properties of the low resolution images to make easier the extraction of heterogeneous patterns from high resolution images, and (2) to enable the analysis of very large images (e.g., 100 000 000 pixels). In order to do so, at each resolution, a collaborative object-based approach is applied to extract the objects of interest. This object-based approach is mainly based on two main steps (i.e., segmentation and classification) which are briefly detailed hereafter.
The segmentation of High Spatial Resolution satellite images is not a trivial task. Indeed, the different objects of interest and thematic ground areas which are sensed by these images, can not be segmented/extracted at the same level (i.e., scale) of segmentation. It is then difficult to correctly segment all these thematic ground areas by using only one segmentation result. In this work, we propose to divide the segmentation step into different sub-steps (i.e., one sub-step per thematic ground area). The segmentation of the images is based on an interactive top-down hierarchical partitioning approach. The proposed algorithm requires a segmentation example of one sample region for each semantic class of the image. Then, the user’s behavior is learnt by using different Machine Learning algorithms and automatically reproduced in the remainder of the image. This process is mainly based on tree-cuts in Binary Partition Trees which are based on Mathematical Morphology definitions.
The classification of the segments is performed by using a multiresolution clustering approach. Instead of characterizing the segments extracted at the current resolution by using features computed on the current image, we propose to characterize these segments by using their decompositions into the next resolution image. To process, for each one of these segments a class-based histogram is computed modeling the composition of this segment in terms of classes into the next resolution image. A classical clustering approach is then performed to create groups of segments sharing similar features. Once these groups have been created, the user can select and recognize them to match with potential thematic ground classes. Distance-based clustering approaches require the definition of a similarity measure to compare the objects which are classified. In this work, we propose a new distance to compare histograms by using semantic and background knowledge provided by the user. The classification results obtained are then embedded into the next resolution image to be traited as input of the segmentation step.
To validate the proposed framework, we have applied it to the urban mapping of complex objects. Experiments have been carried out on multiresolution sets of urban images sensed over different cities. From these datasets, the methodology has been used to extract different levels of details: urban areas from the Low Spatial Resolution images, urban blocks from the High Spatial Resolution images and basic objects from the Very High Spatial Resolution images. The results obtained with this methodology have shown that the quality and the accuracy of the extracted patterns seem sufficient to further accurately perform both classification or object detection in an operational context.
Half thesis manuscript (in French): report half thesis (pdf 8.2 Mb)
Illustration of multiresolution :
 |
 |
 |
Multiresolution set of images |
| 1 pixel = 10 m x 10 m | 1 pixel = 2.4 m x 2.4 m | 1 pixel = 60 cm x 60 cm | |
 |
 |
 |
Groundtruth maps associated |
| Districts | Blocks | Buildings | |
 |
 |
 |
Segmentation / clustering results |
| Districts segmentation | Blocks segmentation | Buildings segmentation |
Research Articles in International Journals (2)
,,Research Articles in International Conference or Workshops Proceedings (5)
,,,,,Research Articles in National Conference or Workshops Proceedings (3)
,Submitted Research Articles (2)
,,Oral communications (5)
Seminars (1)
Posters (1)
,ANR FOSTER (2011 - 2013)
|
The FOSTER project ("Spatio-temporal data mining: application to the understanding and monitoring of soil erosion") is a three-year project launched in January 2011. This computer science project is founded by the French National Research Agency ANR. This project aims at providing to geologists a semi-automatic and complete process for monitoring soil erosion. This process will be based on multi-temporal very high resolution satellite images coupled with digital elevation model (DEM), sensor data and/or expert knowledge. The project will focus more precisely on two important aspects of this process: segmentation of satellite images based on collaborative methods, and construction of descriptive (patterns, clustering, …) and predictive (decision trees, …) spatio-temporal models. |
L1/S1 Physique : Environnements Informatiques
L2/S3 Informatique : Bases de données 1
L2/S4 Informatique : Architecture des ordinateurs
L2/S4 Informatique : Graphes et recherches opérationnelles
L3/S6 : Systèmes distribués
Master (M.Sc.) - Training course
Licence (B.Sc.) - Training course
This html template was designed and provided by Germain Forestier
A link to my parents' website
