Ugo Moschini

Improved detection of faint extended astronomical objects through statistical attribute filtering.

In astronomy, images are produced by sky surveys containing a large number of objects. SExtractor is a widely used program for automated source extraction and cataloguing but struggles with faint extended sources. Using SExtractor as a reference, the paper describes an improvement of a previous method proposed by the authors. It is a Max-Tree-based method for extraction of faint extended sources without stronger image smoothing. Node filtering depends on the noise distribution of a statistic calculated from attributes. Run times are in the same order.

Improving background estimation for faint astronomical object detection

Estimation of the background is an essential step in automated extraction of faint, extended objects from large-scale, optical surveys in astronomy. In this paper we present an improvement on the background estimation method of a commonly used tool in this field: Source Extractor (SEx-tractor). We show that the original method suffers from bias caused by presence of extended sources, and present an alternative which greatly reduces this effect, leading to much better preservation of faint extended structures.

Concurrent computation of connected pattern spectra for very large image information mining

This paper presents a shared-memory parallel algorithm for computing connected pattern spectra from the Max-Tree structure. The pattern spectrum is an aggregated feature space derived directly from the tree-based image representation and is a powerful tool for interactive image information mining. An application example along with timings on experiments with Gpixel input imagery are given. On images of 0.87 to 1.29 Gpixel, wall-clock times of 8.13 to 15.17s, and a speed up of between 27.5 and 33.5 were achieved on a single 2U 64 core rack server.Very-High-Resolution (VHR) images are mapped onto hierarchical segmentations whose segments are characterized by shape and spectral features. Then, new machine learning techniques exploit the resulting tree representation so as to achieve interactive image information mining. The supervised classification of satellite images is an appealing framework for grasping interactively the content of a scene. Existing methods extract spectral, textural and geometric features at the pixel level, before performing a supervised classification. While this approach was shown to provide very good accuracies, the underlying techniques are slow and hardly adaptable thus prohibiting the interactive analysis of scenes. In this paper, we present tree based representations enabling fast information mining from Very High Resolution optical images. VHR images are acquired at resolutions where man-made structures are best associated with homogeneous image regions which are generally captured by a segmentation. The segmentation decreases the number of elements to be managed with respect to the number of pixels. In this paper, the image is mapped onto hierarchical segmentations which have the property of not being constrained by the typical scale of a single segmentation. The image components embedded in the hierarchical segmentations are further characterized by shape and spectral features. The hierarchical nature of the segmentations allows for fast characterization and restitution of the image components. Then, machine learning techniques can be employed for managing the features for performing analysis of the image information content. The paper describes a new machine learning technique which exploits the pre-organization of the features into hierarchical clusterings represented as a tree structure. This machine learning technique is proven to handle millions of elements.ABSTRACTThis paper presents an interactive platform for image infor-mation mining referred to as the switchboard. The switch-board is a dynamic object that interfaces with dedrogramsused for hierarchical image representation. The latter aretree structures used as intermediates between the image spaceand a multi-dimensional feature space inferred from it. Theswitchboard hosts the feature space and facilitates its bi-directional interactivity with the image space through theunderlying tree. Its design ensures operability independent ofthe type of tree it interfaces with. Its functionality is demon-strated using the Alpha-Tree structure for mining buildingfootprints from remote sensing imagery.Index Terms— Alpha-Tree,switchboard,visual-analytics,pattern spectrum, interactive, image information mining.1. INTRODUCTIONImage information mining is often driven by selections ofpositive and negative examples from the image space. Pos-itive examples correspond to manually identified instances ofthe targeted structures and negative to any other undesiredstructure. Examples may be given into the system as tem-plates or regions of interest marked by the user [1]. A pre-specified set of attributes of any such example is mapped intoa feature space and when the latter is populated sufficiently, aclassifier is employed for inferring a decision rule [2]. Basedon this rule the image content can be reduced to only thosestructures that satisfy the imposed criterion.This is a well established paradigm of interactive min-ing which has been optimized through the use of tree struc-tures [1]. A tree is a hierarchical representation of the im-age information content that allows for fast retrieval of imagecomponent attributes. Examples are the Max-Tree [3] andAlpha-Tree [4]. Moreover, tree structures offer the means forcomputing multi-dimensional feature spaces efficiently and

Statistical attribute filtering to detect faint extended astronomical sources

Abstract In astronomy, sky surveys contain a large number of light-emitting sources, often with intensities close to the noise level. Automatic extraction of astronomical objects is therefore needed. SExtractor is a widely used program for automated source extraction and cataloguing, but it is not optimal with faint extended sources. Using SExtractor as a reference, the paper describes an improvement of a previous method proposed by the authors. It is a Max-Tree-based method for extraction of faint extended sources without using a stronger image smoothing. The Max-Tree structure is a hierarchical representation of an image, in which attributes can be computed in every node. Object detection is performed on the nodes of the tree and it relies on the distribution of a statistic calculated using the power attribute, compared to the expected distribution in case of noise. Statistical tests are presented, a comparison with the object extraction of SExtractor is shown and results are discussed.

Viscous-Hyperconnected Attribute Filters: A First Algorithm

In this paper a hyperconnectivity class that tries to address the leakage problem typical of connected filters is used. It shows similarities with the theory of viscous lattices. A novel algorithm to perform attribute filtering of viscous-hyperconnected components is proposed. First, the max-tree of the image eroded by a structuring element is built: it represents the hierarchy of the cores of the hyperconnected components. Then, a processing phase takes place and the node attributes are updated consistently with the pixels of the actual hyperconnected components. Any state-of-the-art algorithm can be used to build the max-tree of the component cores. An issue arises: edges of components are not always correctly preserved. Implementation and performance are presented. A possible solution is put forward and it will be treated in future work.

Parallel 2D Local Pattern Spectra of Invariant Moments for Galaxy Classification

In this paper, we explore the possibility to use 2D pattern spectra as suitable feature vectors in galaxy classification tasks. The focus is on separating mergers from projected galaxies in a data set extracted from the Sloan Digital Sky Survey Data Release 7. Local pattern spectra are built in parallel and are based on an object segmentation obtained by filtering a max-tree structure that preserves faint structures. A set of pattern spectra using size and Hus and Flussers image invariant moments information is computed for every segmented galaxy. The C4.5 tree classifier with bagging gives the best classification result. Mergers and projected galaxies are classified with a precision of about 80%.

ESA-EUSC-JRC 8th Conference on Image Information Mining

This paper presents a shared-memory parallel algorithm for computing connected pattern spectra from the Max-Tree structure. The pattern spectrum is an aggregated feature space derived directly from the tree-based image representation and is a powerful tool for interactive image information mining. An application example along with timings on experiments with Gpixel input imagery are given. On images of 0.87 to 1.29 Gpixel, wall-clock times of 8.13 to 15.17s, and a speed up of between 27.5 and 33.5 were achieved on a single 2U 64 core rack server.Very-High-Resolution (VHR) images are mapped onto hierarchical segmentations whose segments are characterized by shape and spectral features. Then, new machine learning techniques exploit the resulting tree representation so as to achieve interactive image information mining. The supervised classification of satellite images is an appealing framework for grasping interactively the content of a scene. Existing methods extract spectral, textural and geometric features at the pixel level, before performing a supervised classification. While this approach was shown to provide very good accuracies, the underlying techniques are slow and hardly adaptable thus prohibiting the interactive analysis of scenes. In this paper, we present tree based representations enabling fast information mining from Very High Resolution optical images. VHR images are acquired at resolutions where man-made structures are best associated with homogeneous image regions which are generally captured by a segmentation. The segmentation decreases the number of elements to be managed with respect to the number of pixels. In this paper, the image is mapped onto hierarchical segmentations which have the property of not being constrained by the typical scale of a single segmentation. The image components embedded in the hierarchical segmentations are further characterized by shape and spectral features. The hierarchical nature of the segmentations allows for fast characterization and restitution of the image components. Then, machine learning techniques can be employed for managing the features for performing analysis of the image information content. The paper describes a new machine learning technique which exploits the pre-organization of the features into hierarchical clusterings represented as a tree structure. This machine learning technique is proven to handle millions of elements.ABSTRACTThis paper presents an interactive platform for image infor-mation mining referred to as the switchboard. The switch-board is a dynamic object that interfaces with dedrogramsused for hierarchical image representation. The latter aretree structures used as intermediates between the image spaceand a multi-dimensional feature space inferred from it. Theswitchboard hosts the feature space and facilitates its bi-directional interactivity with the image space through theunderlying tree. Its design ensures operability independent ofthe type of tree it interfaces with. Its functionality is demon-strated using the Alpha-Tree structure for mining buildingfootprints from remote sensing imagery.Index Terms— Alpha-Tree,switchboard,visual-analytics,pattern spectrum, interactive, image information mining.1. INTRODUCTIONImage information mining is often driven by selections ofpositive and negative examples from the image space. Pos-itive examples correspond to manually identified instances ofthe targeted structures and negative to any other undesiredstructure. Examples may be given into the system as tem-plates or regions of interest marked by the user [1]. A pre-specified set of attributes of any such example is mapped intoa feature space and when the latter is populated sufficiently, aclassifier is employed for inferring a decision rule [2]. Basedon this rule the image content can be reduced to only thosestructures that satisfy the imposed criterion.This is a well established paradigm of interactive min-ing which has been optimized through the use of tree struc-tures [1]. A tree is a hierarchical representation of the im-age information content that allows for fast retrieval of imagecomponent attributes. Examples are the Max-Tree [3] andAlpha-Tree [4]. Moreover, tree structures offer the means forcomputing multi-dimensional feature spaces efficiently and

Mask Connectivity by Viscous Closings: Linking Merging Galaxies without Merging Double Stars

Second-generation connectivity opened the path to the use of mask images to freely define connectivity among the image components. In theory, any image could be treated as a mask image that defines a certain connectivity. This creates a new problem in terms of which image to use. In this paper, clustering masks suitable for the analysis of astronomical images are discussed. The connectivity defined by such masks must be capable of preserving faint structures like the filaments that link merging galaxies while separating neighboring stars. In this way, the actual morphology of the objects of interest is kept. This is useful for proper segmentation. We show that viscous mathematical morphology operators have a superior performance and create appropriate connectivity masks that can deal with the characteristic features of astronomical images.