Science Service System
You are here : Home : Proposals_Summary

Summary of Proposal MTH0302

TitleModeling and parameter estimation for further classification and recognition tasks, with application to urban/industrial scenes understanding.
Investigator Datcu, Mihai - German Aerospace Center (DLR), Remote Sensing Technology Institute (IMF)
Team Member
PhD Candidate Yao, Wei - Universität Siegen, Zentrum für Sensorsysteme (ZESS)
Mrs. Patrascu, Carmen - University Politehnica Bucharest, Faculty of Electronics and Telecommunications
SummaryObjectives: to study and propose algorithms for information extraction from SSC and detected TerraSAR-X products, such to best characterize their content, i.e. enable to estimate application independednt descriptors for the data which permits in subsequent interpretation process to attain the highest performance in application specific tasks, e.g. image classification, object/structure recognition, quality assessment, indexing, annotation, etc. Methods: In order to capture the whole data complexity, three approaches are proposed: parametric stochastic models for complex valued data, non-parametric statistical analysis, information and complexity theory. We propose an extension of the Gauss-Markov Random Field (GMRF) model for complex-valued data, and for complex-PolSAR data. A next step will be the development of a fully polarimetric multivariate autoregressive process in order to find the complete spatial information delivered by PolSAR data. The comparison of this approach with the information contained in the Wishart distribution will be investigate. Further study and implementation is aiming at non-parametric of eigen-image classifier based on the synergy of two techniques: the Principal Components Analysis (PCA) and the Azimuth sub-band decomposition. Since PCA is describing only linearly dependent structures in the SAR signals, is suitable for analyzing Gaussian data. However, the high resolution SAR images contain edges of different shapes and sizes and could not be described only by Gaussian processes. Thus we combine PCA with Independednt Component analysis ICA. First results have been demonstrated on E-SAR data. The work will continue with the study and elaboration of methods for geometrical and topological description of scenes observed with meter resolution SAR sensors. Adding meaning to images is an important and practical problem in applications as image annotation, indexing, and understanding. Since the meaning is built as function of the meanings of its parts and their mode of syntactic combination, thus is proposed to study how the classical coding theory in relation with the Kolmogorov notion of complexity enables the decomposition of images in an elementary source alphabet, and later using a set of rules to generate a new code with semantic meaning for the image structures. The developed algorithms will be included in the KIM system and aslo used of line to study and asses their performanece and elaborate user specific scenarios for TerraSAR-X data. Also the study will be extended the behaviour of the methods for very large and increasing SAR image volumes, and multi mode sensor data. Asses real applications scenariops and build semantic catalogues for TerraSAR-X archive. Data requirements: Site: any site, which behaves broad variety of objects, and structures, e.g. urban, industrial, rural, vegetation, etc, and also other reference data are available, (can be Munchen, Egypt-the pyramids) Data delivery: Phase 1. For a scene 1, the following products (single polarization): SM SSC, SM MGD, SM GEC, HS SSC, HS MGD, HS GEC. Phase 2. For scene 1 the following products (dual polarization): SM SSC, SM MGD, SM GEC, HS SSC, HS MGD, HS GEC. For scene 2 the following products (single and dual polarization): SM SSC, SM MGD, SM GEC, HS SSC, HS MGD, HS GEC. For one of the scenes the SSC products in single and dual polarization for 4-5 images for look angles aprox. equally spaced in the min.-max. range. Deliverables 1. algorithms for texture and structure parameters estimation for complex and detected data. 2. Algorithms for eigen image decomposition and classification also exploiting the azimuth sublook diversity 3. Algorithms for semantic coding of SAR images 4. Methods and results of evaluation of the previous algorithms in user defined scenarios. Funding: CNES-DLR in the frame of the Competence Center.

Back to list of proposals

© DLR 2004-2016