Skip to Main content Skip to Navigation

Fast learning methods adapted to the user specificities: application to earth observation image information mining

Abstract : An important emerging topic in satellite image content extraction and classification is building retrieval systems that automatically learn high-level semantic interpretations from images, possibly under the direct supervision of the user. Indeed, because of the increased resolution of sensors, the content of satellite images has diversified enormously: it is not uncommon to see details such as cars, streets, technological artefacts, and sometimes people. Thus, retrieval techniques developed initially for multimedia databases are becoming increasingly more relevant to mining data from Earth Observation repositories. The goal of these techniques is to discover new and unexpected patterns, trends, and relationships embedded within large and diverse geographic data sets. Manual annotation is sometimes employed, but is extremely expensive and often subjective. Instead, systems which allow to perform visual content mining from large-scale image databases as well as indexing of high-dimensional database for fast relevant imagery retrieval have been proposed over the past few years. The main contributions of this work are inscribed in the direct continuation of the ideas developed in these systems. We envisage successively the two very broad categories of auto-annotation systems and interactive image search engine to propose our own solutions to the recurring problem of learning from small and non-exhaustive training datasets and of generalizing over a very high-volume of unlabeled data. We first look into the problem of exploiting the huge volume of unlabeled data to discover ''unknown'' semantic structures, that is, semantic classes which are not represented in the training dataset. We among others propose a semi-supervised algorithm able to build an auto-annotation model over non-exhaustive training datasets and to point out to the user new interesting semantic structures in the purpose of guiding him in his database exploration task. In our second contribution, we envisage the problem of speeding up the learning in interactive image search engines. Minimizing the number of iterations in the relevance feedback loop is indeed a crucial issue to build systems which are well-adapted to a human user. With this purpose in mind, we derive a semi-supervised active learning algorithm which exploits the intrinsic data distribution to achieve faster identification of the target category. Our last contribution deals with the problem of retrieving objects in large satellite image scenes. We describe an active learning algorithm which relies on a coarse-to-fine strategy to handle large volumes of data while keeping a satisfying level of accuracy. The proposed algorithm leads to a reduction by more than two orders of magnitude in the number computations necessary at each active learning iteration in standard state-of-the-art interactive image retrieval tools which do not allow to search for complex classes/objects in a really interactive way because of the computational overload inherent to multiple evaluations of the decision function of complex classifiers. We assess each time our results on Spot5 and QuickBird panchromatic imagery and we show that the methods we propose significantly outperform state-of-the-art techniques while adding interesting new features such as the "unknown'' semantic structures discovery feature in the auto-annotation case or the interactive search scheme in the object retrieval part.
Document type :
Complete list of metadata
Contributor : Pierre Blanchart Connect in order to contact the contributor
Submitted on : Wednesday, January 25, 2012 - 9:28:03 AM
Last modification on : Friday, July 31, 2020 - 10:44:07 AM
Long-term archiving on: : Wednesday, December 14, 2016 - 12:42:31 AM


  • HAL Id : pastel-00662747, version 1



Pierre Blanchart. Fast learning methods adapted to the user specificities: application to earth observation image information mining. Engineering Sciences [physics]. Télécom ParisTech, 2011. English. ⟨pastel-00662747⟩



Record views


Files downloads