Papers by Sébastien Lefèvre
Publication in the conference proceedings of EUSIPCO, Florence, Italy, 2006
2017 Joint Urban Remote Sensing Event (JURSE), 2017
This work shows how deep learning techniques can benefit to remote sensing. We focus on tasks whi... more This work shows how deep learning techniques can benefit to remote sensing. We focus on tasks which are recurrent in Earth Observation data analysis. For classification and semantic mapping of aerial images, we present various deep network architectures and show that context information and dense labeling allow to reach better performances. For estimation of normals in point clouds, combining Hough transform with convolutional networks also improves the accuracy of previous frameworks by detecting hard configurations like corners. It shows that deep learning allows to revisit remote sensing and offers promising paths for urban modeling and monitoring.
2014 6th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2014
Mapping of remote sensing data is usually done through image classification. For hyperspectral im... more Mapping of remote sensing data is usually done through image classification. For hyperspectral images, the classification process often relies only on the spectral signature of each single pixel. Nevertheless, combining spatial and spectral features has been a promising way for accuracy improvement. We address here this problem by computing spectral features from spatially identified regions, sampled from a hierarchical image representation, namely α-tree, built with prior knowledge. The sampling of the tree nodes (i.e., regions) is based on the paradigm of constrained connectivity and the global range criterion. In this paper, we extend this criterion to hyperspectral data and apply it to our knowledge-based α-tree. Our results show an improvement of pixelwise classification accuracy over spectral features only.
ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2020
Monitoring observable processes in Satellite Image Time Series (SITS) is one of the crucial way t... more Monitoring observable processes in Satellite Image Time Series (SITS) is one of the crucial way to understand dynamics of our planet that is facing unexpected behaviors due to climate change. In this paper, we propose a novel method to assess the evolution of objects (and especially their surface) through time. To do so, we first build a space-time tree representation of image time series. The so-called space-time tree is a hierarchical representation of an image sequences into a nested set of nodes characterizing the observed regions at multiple spatial and temporal scales. Then, we measure for each node the spatial area occupied at each time sample, and we focus on its evolution through time. We thus define the spatio-temporal stability of each node. We use this attribute to identify and measure changing areas in a remotely-sensed scene. We illustrate the purpose of our method with some experiments in a coastal environment using Sentinel-2 images, and in a flood occurred area with Sentinel-1 images.
Computer Vision and Image Understanding, 2019
Understanding visual scenes relies more and more on dense pixel-wise classification obtained via ... more Understanding visual scenes relies more and more on dense pixel-wise classification obtained via deep fully convolutional neural networks. However, due to the nature of the networks, predictions often suffer from blurry boundaries and ill-segmented shapes, fueling the need for post-processing. This work introduces a new semantic segmentation regularization based on the regression of a distance transform. After computing the distance transform on the label masks, we train a FCN in a multi-task setting in both discrete and continuous spaces by learning jointly classification and distance regression. This requires almost no modification of the network structure and adds a very low overhead to the training process. Learning to approximate the distance transform back-propagates spatial cues that implicitly regularizes the segmentation. We validate this technique with several architectures on various datasets, and we show significant improvements compared to competitive baselines.
Pattern Recognition, 2009
The morphological Hit-or-Miss Transform (HMT) is a powerful tool for digital image analysis. Its ... more The morphological Hit-or-Miss Transform (HMT) is a powerful tool for digital image analysis. Its recent extensions to grey level images have proven its ability to solve various template matching problems. In this paper we explore the capacity of various existing approaches to work in very noisy environments and discuss the generic methods used to improve their robustness to noise. We also propose a new formulation for a fuzzy morphological HMT which has been especially designed to deal with very noisy images. Our approach is validated through a pattern matching problem in astronomical images that consists of detecting very faint objects: low surface brightness galaxies. Despite their influence on the galactic evolution model, these objects remain mostly misunderstood by the astronomers. Due to their low signal to noise ratio, there is no automatic and reliable detection method yet. In this paper we introduce such a method based on the proposed hit-or-miss operator. The complete process is described starting from the building of a set of patterns until the reconstruction of a suitable map of detected objects. Implementation, running cost and optimisations are discussed. Outcomes have been examined by astronomers and compared to previous works. We have observed promising results in this difficult context for which Mathemat
ISPRS Journal of Photogrammetry and Remote Sensing, 2018
In this work, we investigate various methods to deal with semantic labeling of very high resoluti... more In this work, we investigate various methods to deal with semantic labeling of very high resolution multi-modal remote sensing data. Especially, we study how deep fully convolutional networks can be adapted to deal with multi-modal and multi-scale remote sensing data for semantic labeling. Our contributions are threefold: a) we present an efficient multi-scale approach to leverage both a large spatial context and the high resolution data, b) we investigate early and late fusion of Lidar and multispectral data, c) we validate our methods on two public datasets with state-of-the-art results. Our results indicate that late fusion make it possible to recover errors steaming from ambiguous data, while early fusion allows for better joint-feature learning but at the cost of higher sensitivity to missing data.
Pattern Recognition, 2007
The successful application of univariate morphological operators on several domains, along with t... more The successful application of univariate morphological operators on several domains, along with the increasing need for processing the plethora of available multivalued images, have been the main motives behind the efforts concentrated on extending the mathematical morphology framework to multivariate data. The few theoretical requirements of this extension, consisting primarily of a ranking scheme as well as extrema operators for vectorial data, have led to numerous suggestions with diverse properties. However, none of them has yet been widely accepted. Furthermore, the comparison research work in the current literature, evaluating the results obtained from these approaches, is either outdated or limited to a particular application domain. In this paper, a comprehensive review of the proposed multivariate morphological frameworks is provided. In particular, they are examined mainly with respect to their data ordering methodologies. Additionally, the results of a brief series of illustrative application oriented tests of selected vector orderings on colour and multispectral remote sensing data are also discussed.
Lecture Notes in Computer Science
Due to its broad impact in many image analysis applications, the problem of image segmentation ha... more Due to its broad impact in many image analysis applications, the problem of image segmentation has been widely studied. However, there still does not exist any automatic segmentation procedure able to deal accurately with any kind of image. Thus semi-automatic segmentation methods may be seen as an appropriate alternative to solve the segmentation problem. Among these methods, the marker-based watershed has been successfully involved in various domains. In this algorithm, the user may locate the markers, which are used only as the initial starting positions of the regions to be segmented. We propose to base the segmentation process also on the contents of the markers through a supervised pixel classification, thus resulting in a knowledge-based watershed segmentation where the knowledge is built from the markers. Our contribution has been evaluated through some comparative tests with some state-of-the-art methods on the well-known Berkeley Segmentation Dataset.
Lecture Notes in Computer Science, 2008
The Hit-or-Miss transform is a well-known morphological operator for template matching in binary ... more The Hit-or-Miss transform is a well-known morphological operator for template matching in binary and grey-level images. However it cannot be used straightforward in multivalued images (such as colour or multispectral images) since Mathematical Morphology needs an ordering relation which is not trivial on multivalued spaces. Moreover, existing definitions of the Hit-Or-Miss Transform in grey-level use only spatial templates (or structuring elements) which could be insufficient for some feature extraction problems. In this paper, we propose a multivariate Hitor-Miss Transform operator which combines spatial and spectral patterns to perform template matching. We illustrate its relevance with an application in the remote sensing field, the extraction of coastline from very high (spatial) resolution images.
Lecture Notes in Computer Science, 2011
In order to face the various needs of users, user-driven segmentation methods are expected to pro... more In order to face the various needs of users, user-driven segmentation methods are expected to provide more relevant results than fully automatic approaches. Within Mathematical Morphology, several user-driven approaches have been proposed, mostly relying on the watershed transform. Nevertheless, Soille (IEEE TPAMI, 2008) has recently suggested another solution by gathering puzzle pieces computed as Quasi-Flat Zones (QFZ) of an image. In this paper, we study more deeply this user-driven segmentation scheme in the context of video data. Thus we also introduce the concept of Spatio-Temporal QFZ and propose several methods for extracting such zones from a video sequence.
Eighth International Conference on Document Analysis and Recognition (ICDAR'05), 2005
In this article, we focus on the problem of caption detection in video sequences. Contrary to mos... more In this article, we focus on the problem of caption detection in video sequences. Contrary to most of existing approaches based on a single detector followed by an ad hoc and costly post-processing, we have decided to consider several detectors and to merge their results in order to combine advantages of each one. First we made a study of captions in video sequences to determine how they are represented in images and to identify their main features (color constancy and background contrast, edge density and regularity, temporal persistence). Based on these features, we then select or define the appropriate detectors and we compare several fusion strategies which can be involved. The logical process we have followed and the satisfying results we have obtained let us validate our contribution.
2008 IAPR Workshop on Pattern Recognition in Remote Sensing (PRRS 2008), 2008
This paper presents the initial results of the Algorithm Performance Contest that was organized a... more This paper presents the initial results of the Algorithm Performance Contest that was organized as part of the 5th IAPR Workshop on Pattern Recognition in Remote Sensing (PRRS 2008). The focus of the 2008 contest was automatic building detection and digital surface model (DSM) extraction. A QuickBird data set with manual ground truth was used for building detection evaluation, and a stereo Ikonos data set with a highly accurate reference DSM was used for DSM extraction evaluation. Nine submissions were received for the building detection task, and three submissions were received for the DSM extraction task. We provide an overview of the data sets, the summaries of the methods used for the submissions, the details of the evaluation criteria, and the results of the initial evaluation.
Journal of Real-Time Image Processing, 2007
In this article, we deal with the problem of shot change detection which is of primary importance... more In this article, we deal with the problem of shot change detection which is of primary importance when trying to segment and abstract video sequences. Contrary to recent experiments, our aim is to elaborate a robust but very efficient (real-time even with uncompressed data) method to deal with the remaining problems related to shot change detection: illumination changes, context and data independency, and parameter settings. To do so, we have considered some adaptive threshold and derivative measures in a hue-saturation colour space. We illustrate our robust and efficient method by some experiments on news and football broadcast video sequences.
Revue internationale de géomatique, 2007
Cet article présente une nouvelle méthode de détection et d'extraction des bâtiments en milieu ur... more Cet article présente une nouvelle méthode de détection et d'extraction des bâtiments en milieu urbain à partir d'images satellitaires à très haute résolution spatiale. L'approche proposée est fondée sur l'application et l'enchainement automatique d'opérateurs issus de la morphologie mathématique binaire. Plusieurs étapes constituent la méthode : (1) binarisation de l'image, (2) filtrage du bruit et des éléments de taille inférieure aux bâtiments par lissage morphologique, (3) détection des bâtiments par application d'une transformée en tout ou rien adaptative, avec un élément structurant de taille et de forme variable, (4) restauration de la forme des bâtiments par reconstruction géodésique. Deux stratégies différentes de binarisation sont proposées lors de l'étape initiale. La première consiste à binariser l'image par seuillage, le seuil étant défini soit de manière automatique, soit de manière empirique en fonction de l'image traitée. La seconde est fondée sur l'application d'une classification non supervisée pour laquelle le nombre de classes n'est pas fixé a priori. La méthode a été mise en oeuvre sur une image Quickbird panchromatique de la région de Strasbourg. Les résultats obtenus confirment l'intérêt et l'efficacité de l'approche. ABSTRACT. This paper presents a new method for building extraction in Very High Resolution remotely sensed images in urban areas. The approach proposed is based on the use binary mathematical morphology operators. The method is composed of several steps: (1) conversion of grey level images to binary images, (2) smoothing by means of morphological filtering, (3) building detection with an adaptive hit-or-miss transform, (4) shape restoration. Two strategies of binarization are proposed. The first one consists in performing an interactive or automatic thresholding. The second one is based on an unsupervised classification. The method has been applied on a Quickbird panchromatic image. Results show the interest of the approach.
2007 Urban Remote Sensing Joint Event, 2007
This paper presents a new method for buildings extraction in Very High Resolution (VHR) remotely ... more This paper presents a new method for buildings extraction in Very High Resolution (VHR) remotely sensed images based on binary mathematical morphology (MM) operators. The proposed approach involves several advanced morphological operators among which an adaptive hit-or-miss transform with varying sizes and shapes of the structuring element and a bidimensional granulometry intended to determine the optimal filtering parameters automatically. A clustering-based approach for image binarization is also introduced. This one avoids an empirical thresholding of input panchromatic images. Experiments made on a Quickbird VHR-image show the effectiveness of the method.
Journal of pathology informatics, 2013
In the framework of the Cognitive Microscope (MICO) project, we have set up a contest about mitos... more In the framework of the Cognitive Microscope (MICO) project, we have set up a contest about mitosis detection in images of H and E stained slides of breast cancer for the conference ICPR 2012. Mitotic count is an important parameter for the prognosis of breast cancer. However, mitosis detection in digital histopathology is a challenging problem that needs a deeper study. Indeed, mitosis detection is difficult because mitosis are small objects with a large variety of shapes, and they can thus be easily confused with some other objects or artefacts present in the image. We added a further dimension to the contest by using two different slide scanners having different resolutions and producing red-green-blue (RGB) images, and a multi-spectral microscope producing images in 10 different spectral bands and 17 layers Z-stack. 17 teams participated in the study and the best team achieved a recall rate of 0.7 and precision of 0.89. Several studies on automatic tools to process digitized sli...
Journal of Electronic Imaging, 2009
In the field of digital image processing, the description of image content is one of the most cru... more In the field of digital image processing, the description of image content is one of the most crucial tasks. Indeed, it is a mandatory step for various applications, such as industrial vision, medical imaging, content-based image retrieval, etc. The description of the image content is achieved through the computation of some predefined features, which can be performed at different scales. Among global features that describe the content of the whole image, the gray level histogram focuses on the distribution of gray levels within the image, while morphological features (e.g., the pattern spectrum) measure the distribution of object sizes in the image. Despite their broad interest, such morphological size-distribution features are limited due to their monodimensional nature. Our goal is to review multidimensional extensions of these features able to deal with complementary information (such as shape, orientation, spectral, intensity, or spatial information). Moreover, we illustrate each multidimensional feature by an illustrative example that shows their relevance compared to the standard morphological size distribution. These features can be seen as relevant solutions when the standard monodimensional features fail to accurately represent the image content.
dpt-info.u-strasbg.fr
Résumé Dans cet article, nous considérons les séquences vidéo couleur comme des données complexes... more Résumé Dans cet article, nous considérons les séquences vidéo couleur comme des données complexes. Notre contribution porte sur deux méthodes adaptées à ce type de données et dont l'objectif est d'extraire respectivement des indices spatiaux et temporels. Nous pensons que ces méthodes d'extraction d'indices peuvent être intégrées avec succès dans un processus plus complexe de fouille de données multimédia, aspect qui ne sera pas abordé ici. Les méthodes que nous présentons ici sont basées sur l'espace Teinte ...
Résumé. L'apprentissage d'algorithmes d'interprétation d'images est un proces... more Résumé. L'apprentissage d'algorithmes d'interprétation d'images est un processus complexe de fouille de données. La méthode consistant à considérer les pixels de façon indépendante a montré ses limites. En effet, les classes d'intérêt ne sont pas toujours séparables en utilisant uniquement les caractéristiques relatives aux pixels. Pour pallier à ce problème, les méthodes actuelles d'interprétation d'images s' appuient sur une segmentation préalable de l'image qui consiste en une agrégation des pixels connexes ...
Uploads
Papers by Sébastien Lefèvre