Advanced Kernel Methods For Remote Sensing Image Classification

Résumé Suite aux recentes avancées technologiques, les archives d'images digitales ont connu une croissance qualitative et quantitative sans précédent. Malgré les énormes possibilités qu'elles offrent, ces avancées posent de nouvelles questions quant au traitement des masses de données saisies. Cette question est à la base de cette Thèse: les problèmes de traitement d'information digitale à très haute résolution spatiale et/ou spectrale y sont considérés en recourant à des approches d'apprentissage statistique, les méthodes à noyau. Cette Thèse étudie des problèmes de classification d'images, c'est à dire de catégorisation de pixels en un nombre réduit de classes refletant les propriétés spectrales et contextuelles des objets qu'elles représentent. L'accent est mis sur l'efficience des algorithmes, ainsi que sur leur simplicité, de manière à augmenter leur potentiel d'implementation pour les utilisateurs. De plus, le défi de cette Thèse est de rester proche des problèmes concrets des utilisateurs d'images satellite sans pour autant perdre de vue l'intéret des méthodes proposées pour le milieu du machine learning dont elles sont issues. En ce sens, ce travail joue la carte de la transdisciplinarité en maintenant un lien fort entre les deux sciences dans tous les développements proposés. Quatre modèles sont proposés: le premier répond au problème de la haute dimensionalité et de la redondance des données par un modèle optimisant les performances en classification en s'adaptant aux particularités de l'image. Ceci est rendu possible par un système de ranking des variables (les bandes) qui est optimisé en même temps que le modèle de base: ce faisant, seules les variables importantes pour résoudre le problème sont utilisées par le classifieur. Le manque d'information étiquétée et l'incertitude quant à sa pertinence pour le problème sont à la source des deux modèles suivants, basés respectivement sur l'apprentissage actif et les méthodes semi-supervisées: le premier permet d'améliorer la qualité d'un ensemble d'entraînement par interaction directe entre l'utilisateur et la machine, alors que le deuxième utilise les pixels non étiquetés pour améliorer la description des données disponibles et la robustesse du modèle. Enfin, le dernier modèle proposé considère la question plus théorique de la structure entre les outputs: l'intègration de cette source d'information, jusqu'à présent jamais considérée en télédétection, ouvre des nouveaux défis de recherche. Advanced kernel methods for remote sensing image classification Devis Tuia Institut de Géomatique et d'Analyse du Risque September 2009 Abstract The technical developments in recent years have brought the quantity and quality of digital information to an unprecedented level, as enormous archives of satellite images are available to the users. However, even if these advances open more and more possibilities in the use of digital imagery, they also rise several problems of storage and treatment. The latter is considered in this Thesis: the processing of very high spatial and spectral resolution images is treated with approaches based on data-driven algorithms relying on kernel methods. In particular, the problem of image classification, i.e. the categorization of the image's pixels into a reduced number of classes reflecting spectral and contextual properties, is studied through the different models presented. The accent is put on algorithmic efficiency and the simplicity of the approaches proposed, to avoid too complex models that would not be used by users. The major challenge of the Thesis is to remain close to concrete remote sensing problems, without losing the methodological interest from the machine learning viewpoint: in this sense, this work aims at building a bridge between the machine learning and remote sensing communities and all the models proposed have been developed keeping in mind the need for such a synergy. Four models are proposed: first, an adaptive model learning the relevant image features has been proposed to solve the problem of high dimensionality and collinearity of the image features. This model provides automatically an accurate classifier and a ranking of the relevance of the single features. The scarcity and unreliability of labeled. information were the common root of the second and third models proposed: when confronted to such problems, the user can either construct the labeled set iteratively by direct interaction with the machine or use the unlabeled data to increase robustness and quality of the description of data. Both solutions have been explored resulting into two methodological contributions, based respectively on active learning and semisupervised learning. Finally, the more theoretical issue of structured outputs has been considered in the last model, which, by integrating outputs similarity into a model, opens new challenges and opportunities for remote sensing image processing.

Objective assessment of image quality in conventional and digital mammography taking into account dynamic range.

The goal of this work is to develop a method to objectively compare the performance of a digital and a screen-film mammography system in terms of image quality. The method takes into account the dynamic range of the image detector, the detection of high and low contrast structures, the visualisation of the images and the observer response. A test object, designed to represent a compressed breast, was constructed from various tissue equivalent materials ranging from purely adipose to purely glandular composition. Different areas within the test object permitted the evaluation of low and high contrast detection, spatial resolution and image noise. All the images (digital and conventional) were captured using a CCD camera to include the visualisation process in the image quality assessment. A mathematical model observer (non-prewhitening matched filter), that calculates the detectability of high and low contrast structures using spatial resolution, noise and contrast, was used to compare the two technologies. Our results show that for a given patient dose, the detection of high and low contrast structures is significantly better for the digital system than for the conventional screen-film system studied. The method of using a test object with a large tissue composition range combined with a camera to compare conventional and digital imaging modalities can be applied to other radiological imaging techniques. In particular it could be used to optimise the process of radiographic reading of soft copy images.

Fast Geodesic Active Fields for Image Registration Based on Splitting and Augmented Lagrangian Approaches.

In this paper, we present an efficient numerical scheme for the recently introduced geodesic active fields (GAF) framework for geometric image registration. This framework considers the registration task as a weighted minimal surface problem. Hence, the data-term and the regularization-term are combined through multiplication in a single, parametrization invariant and geometric cost functional. The multiplicative coupling provides an intrinsic, spatially varying and data-dependent tuning of the regularization strength, and the parametrization invariance allows working with images of nonflat geometry, generally defined on any smoothly parametrizable manifold. The resulting energy-minimizing flow, however, has poor numerical properties. Here, we provide an efficient numerical scheme that uses a splitting approach; data and regularity terms are optimized over two distinct deformation fields that are constrained to be equal via an augmented Lagrangian approach. Our approach is more flexible than standard Gaussian regularization, since one can interpolate freely between isotropic Gaussian and anisotropic TV-like smoothing. In this paper, we compare the geodesic active fields method with the popular Demons method and three more recent state-of-the-art algorithms: NL-optical flow, MRF image registration, and landmark-enhanced large displacement optical flow. Thus, we can show the advantages of the proposed FastGAF method. It compares favorably against Demons, both in terms of registration speed and quality. Over the range of example applications, it also consistently produces results not far from more dedicated state-of-the-art methods, illustrating the flexibility of the proposed framework.

Model-based Segmentation and Image Fusion of 3D Computed Tomography and 3D Ultrasound of the Eye for Radiotherapy Planning

For radiotherapy treatment planning of retinoblastoma inchildhood, Computed Tomography (CT) represents thestandard method for tumor volume delineation, despitesome inherent limitations. CT scan is very useful inproviding information on physical density for dosecalculation and morphological volumetric information butpresents a low sensitivity in assessing the tumorviability. On the other hand, 3D ultrasound (US) allows ahigh accurate definition of the tumor volume thanks toits high spatial resolution but it is not currentlyintegrated in the treatment planning but used only fordiagnosis and follow-up. Our ultimate goal is anautomatic segmentation of gross tumor volume (GTV) in the3D US, the segmentation of the organs at risk (OAR) inthe CT and the registration of both. In this paper, wepresent some preliminary results in this direction. Wepresent 3D active contour-based segmentation of the eyeball and the lens in CT images; the presented approachincorporates the prior knowledge of the anatomy by usinga 3D geometrical eye model. The automated segmentationresults are validated by comparing with manualsegmentations. Then, for the fusion of 3D CT and USimages, we present two approaches: (i) landmark-basedtransformation, and (ii) object-based transformation thatmakes use of eye ball contour information on CT and USimages.

Free-breathing whole-heart coronary MRA with 3D radial SSFP and self-navigated image reconstruction.

Respiratory motion is a major source of artifacts in cardiac magnetic resonance imaging (MRI). Free-breathing techniques with pencil-beam navigators efficiently suppress respiratory motion and minimize the need for patient cooperation. However, the correlation between the measured navigator position and the actual position of the heart may be adversely affected by hysteretic effects, navigator position, and temporal delays between the navigators and the image acquisition. In addition, irregular breathing patterns during navigator-gated scanning may result in low scan efficiency and prolonged scan time. The purpose of this study was to develop and implement a self-navigated, free-breathing, whole-heart 3D coronary MRI technique that would overcome these shortcomings and improve the ease-of-use of coronary MRI. A signal synchronous with respiration was extracted directly from the echoes acquired for imaging, and the motion information was used for retrospective, rigid-body, through-plane motion correction. The images obtained from the self-navigated reconstruction were compared with the results from conventional, prospective, pencil-beam navigator tracking. Image quality was improved in phantom studies using self-navigation, while equivalent results were obtained with both techniques in preliminary in vivo studies.

An Image Enhancement Technique using MTF of Human eye

Image filtering is a highly demanded approach of image enhancement in digital imaging systems design. It is widely used in television and camera design technologies to improve the quality of an output image to avoid various problems such as image blurring problem thatgains importance in design of displays of large sizes and design of digital cameras. This thesis proposes a new image filtering method basedon visual characteristics of human eye such as MTF. In contrast to the traditional filtering methods based on human visual characteristics this thesis takes into account the anisotropy of the human eye vision. The proposed method is based on laboratory measurements of the human eye MTF and takes into account degradation of the image by the latter. This method improves an image in the way it will be degraded by human eye MTF to give perception of the original image quality. This thesis gives a basic understanding of an image filtering approach and the concept of MTF and describes an algorithm to perform an image enhancement based on MTF of human eye. Performed experiments have shown quite good results according to human evaluation. Suggestions to improve the algorithm are also given for the future improvements.

Sparse image reconstruction on the sphere: implications of a new sampling theorem.

We study the impact of sampling theorems on the fidelity of sparse image reconstruction on the sphere. We discuss how a reduction in the number of samples required to represent all information content of a band-limited signal acts to improve the fidelity of sparse image reconstruction, through both the dimensionality and sparsity of signals. To demonstrate this result, we consider a simple inpainting problem on the sphere and consider images sparse in the magnitude of their gradient. We develop a framework for total variation inpainting on the sphere, including fast methods to render the inpainting problem computationally feasible at high resolution. Recently a new sampling theorem on the sphere was developed, reducing the required number of samples by a factor of two for equiangular sampling schemes. Through numerical simulations, we verify the enhanced fidelity of sparse image reconstruction due to the more efficient sampling of the sphere provided by the new sampling theorem.

Retinal Image Analysis Using Machine Vision

The topic of this thesis is studying how lesions in retina caused by diabetic retinopathy can be detected from color fundus images by using machine vision methods. Methods for equalizing uneven illumination in fundus images, detecting regions of poor image quality due toinadequate illumination, and recognizing abnormal lesions were developed duringthe work. The developed methods exploit mainly the color information and simpleshape features to detect lesions. In addition, a graphical tool for collecting lesion data was developed. The tool was used by an ophthalmologist who marked lesions in the images to help method development and evaluation. The tool is a general purpose one, and thus it is possible to reuse the tool in similar projects.The developed methods were tested with a separate test set of 128 color fundus images. From test results it was calculated how accurately methods classify abnormal funduses as abnormal (sensitivity) and healthy funduses as normal (specificity). The sensitivity values were 92% for hemorrhages, 73% for red small dots (microaneurysms and small hemorrhages), and 77% for exudates (hard and soft exudates). The specificity values were 75% for hemorrhages, 70% for red small dots, and 50% for exudates. Thus, the developed methods detected hemorrhages accurately and microaneurysms and exudates moderately.

Brain surface non-rigid registration in auditory processing

Résumé: Le développement rapide de nouvelles technologies comme l'imagerie médicale a permis l'expansion des études sur les fonctions cérébrales. Le rôle principal des études fonctionnelles cérébrales est de comparer l'activation neuronale entre différents individus. Dans ce contexte, la variabilité anatomique de la taille et de la forme du cerveau pose un problème majeur. Les méthodes actuelles permettent les comparaisons interindividuelles par la normalisation des cerveaux en utilisant un cerveau standard. Les cerveaux standards les plus utilisés actuellement sont le cerveau de Talairach et le cerveau de l'Institut Neurologique de Montréal (MNI) (SPM99). Les méthodes de recalage qui utilisent le cerveau de Talairach, ou celui de MNI, ne sont pas suffisamment précises pour superposer les parties plus variables d'un cortex cérébral (p.ex., le néocortex ou la zone perisylvienne), ainsi que les régions qui ont une asymétrie très importante entre les deux hémisphères. Le but de ce projet est d'évaluer une nouvelle technique de traitement d'images basée sur le recalage non-rigide et utilisant les repères anatomiques. Tout d'abord, nous devons identifier et extraire les structures anatomiques (les repères anatomiques) dans le cerveau à déformer et celui de référence. La correspondance entre ces deux jeux de repères nous permet de déterminer en 3D la déformation appropriée. Pour les repères anatomiques, nous utilisons six points de contrôle qui sont situés : un sur le gyrus de Heschl, un sur la zone motrice de la main et le dernier sur la fissure sylvienne, bilatéralement. Evaluation de notre programme de recalage est accomplie sur les images d'IRM et d'IRMf de neuf sujets parmi dix-huit qui ont participés dans une étude précédente de Maeder et al. Le résultat sur les images anatomiques, IRM, montre le déplacement des repères anatomiques du cerveau à déformer à la position des repères anatomiques de cerveau de référence. La distance du cerveau à déformer par rapport au cerveau de référence diminue après le recalage. Le recalage des images fonctionnelles, IRMf, ne montre pas de variation significative. Le petit nombre de repères, six points de contrôle, n'est pas suffisant pour produire les modifications des cartes statistiques. Cette thèse ouvre la voie à une nouvelle technique de recalage du cortex cérébral dont la direction principale est le recalage de plusieurs points représentant un sillon cérébral. Abstract : The fast development of new technologies such as digital medical imaging brought to the expansion of brain functional studies. One of the methodolgical key issue in brain functional studies is to compare neuronal activation between individuals. In this context, the great variability of brain size and shape is a major problem. Current methods allow inter-individual comparisions by means of normalisation of subjects' brains in relation to a standard brain. A largerly used standard brains are the proportional grid of Talairach and Tournoux and the Montreal Neurological Insititute standard brain (SPM99). However, there is a lack of more precise methods for the superposition of more variable portions of the cerebral cortex (e.g, neocrotex and perisyvlian zone) and in brain regions highly asymmetric between the two cerebral hemipsheres (e.g. planum termporale). The aim of this thesis is to evaluate a new image processing technique based on non-linear model-based registration. Contrary to the intensity-based, model-based registration uses spatial and not intensitiy information to fit one image to another. We extract identifiable anatomical features (point landmarks) in both deforming and target images and by their correspondence we determine the appropriate deformation in 3D. As landmarks, we use six control points that are situated: one on the Heschl'y Gyrus, one on the motor hand area, and one on the sylvian fissure, bilaterally. The evaluation of this model-based approach is performed on MRI and fMRI images of nine of eighteen subjects participating in the Maeder et al. study. Results on anatomical, i.e. MRI, images, show the mouvement of the deforming brain control points to the location of the reference brain control points. The distance of the deforming brain to the reference brain is smallest after the registration compared to the distance before the registration. Registration of functional images, i.e fMRI, doesn't show a significant variation. The small number of registration landmarks, i.e. six, is obvious not sufficient to produce significant modification on the fMRI statistical maps. This thesis opens the way to a new computation technique for cortex registration in which the main directions will be improvement of the registation algorithm, using not only one point as landmark, but many points, representing one particular sulcus.

Preferred Spatial Frequencies for Human Face Processing Are Associated with Optimal Class Discrimination in the Machine

Psychophysical studies suggest that humans preferentially use a narrow band of low spatial frequencies for face recognition. Here we asked whether artificial face recognition systems have an improved recognition performance at the same spatial frequencies as humans. To this end, we estimated recognition performance over a large database of face images by computing three discriminability measures: Fisher Linear Discriminant Analysis, Non-Parametric Discriminant Analysis, and Mutual Information. In order to address frequency dependence, discriminabilities were measured as a function of (filtered) image size. All three measures revealed a maximum at the same image sizes, where the spatial frequency content corresponds to the psychophysical found frequencies. Our results therefore support the notion that the critical band of spatial frequencies for face recognition in humans and machines follows from inherent properties of face images, and that the use of these frequencies is associated with optimal face recognition performance.

Environmental Monitoring using Image Analysis

Tärkeä tehtävä ympäristön tarkkailussa on arvioida ympäristön nykyinen tila ja ihmisen siihen aiheuttamat muutokset sekä analysoida ja etsiä näiden yhtenäiset suhteet. Ympäristön muuttumista voidaan hallita keräämällä ja analysoimalla tietoa. Tässä diplomityössä on tutkittu vesikasvillisuudessa hai vainuja muutoksia käyttäen etäältä hankittua mittausdataa ja kuvan analysointimenetelmiä. Ympäristön tarkkailuun on käytetty Suomen suurimmasta järvestä Saimaasta vuosina 1996 ja 1999 otettuja ilmakuvia. Ensimmäinen kuva-analyysin vaihe on geometrinen korjaus, jonka tarkoituksena on kohdistaa ja suhteuttaa otetut kuvat samaan koordinaattijärjestelmään. Toinen vaihe on kohdistaa vastaavat paikalliset alueet ja tunnistaa kasvillisuuden muuttuminen. Kasvillisuuden tunnistamiseen on käytetty erilaisia lähestymistapoja sisältäen valvottuja ja valvomattomia tunnistustapoja. Tutkimuksessa käytettiin aitoa, kohinoista mittausdataa, minkä perusteella tehdyt kokeet antoivat hyviä tuloksia tutkimuksen onnistumisesta.

Unsupervised active regions for multiresolution image segmentation

An unsupervised approach to image segmentation which fuses region and boundary information is presented. The proposed approach takes advantage of the combined use of 3 different strategies: the guidance of seed placement, the control of decision criterion, and the boundary refinement. The new algorithm uses the boundary information to initialize a set of active regions which compete for the pixels in order to segment the whole image. The method is implemented on a multiresolution representation which ensures noise robustness as well as computation efficiency. The accuracy of the segmentation results has been proven through an objective comparative evaluation of the method

New Distance Transforms for Gray Level Image Compression

This thesis deals with distance transforms which are a fundamental issue in image processing and computer vision. In this thesis, two new distance transforms for gray level images are presented. As a new application for distance transforms, they are applied to gray level image compression. The new distance transforms are both new extensions of the well known distance transform algorithm developed by Rosenfeld, Pfaltz and Lay. With some modification their algorithm which calculates a distance transform on binary images with a chosen kernel has been made to calculate a chessboard like distance transform with integer numbers (DTOCS) and a real value distance transform (EDTOCS) on gray level images. Both distance transforms, the DTOCS and EDTOCS, require only two passes over the graylevel image and are extremely simple to implement. Only two image buffers are needed: The original gray level image and the binary image which defines the region(s) of calculation. No other image buffers are needed even if more than one iteration round is performed. For large neighborhoods and complicated images the two pass distance algorithm has to be applied to the image more than once, typically 3 10 times. Different types of kernels can be adopted. It is important to notice that no other existing transform calculates the same kind of distance map as the DTOCS. All the other gray weighted distance function, GRAYMAT etc. algorithms find the minimum path joining two points by the smallest sum of gray levels or weighting the distance values directly by the gray levels in some manner. The DTOCS does not weight them that way. The DTOCS gives a weighted version of the chessboard distance map. The weights are not constant, but gray value differences of the original image. The difference between the DTOCS map and other distance transforms for gray level images is shown. The difference between the DTOCS and EDTOCS is that the EDTOCS calculates these gray level differences in a different way. It propagates local Euclidean distances inside a kernel. Analytical derivations of some results concerning the DTOCS and the EDTOCS are presented. Commonly distance transforms are used for feature extraction in pattern recognition and learning. Their use in image compression is very rare. This thesis introduces a new application area for distance transforms. Three new image compression algorithms based on the DTOCS and one based on the EDTOCS are presented. Control points, i.e. points that are considered fundamental for the reconstruction of the image, are selected from the gray level image using the DTOCS and the EDTOCS. The first group of methods select the maximas of the distance image to new control points and the second group of methods compare the DTOCS distance to binary image chessboard distance. The effect of applying threshold masks of different sizes along the threshold boundaries is studied. The time complexity of the compression algorithms is analyzed both analytically and experimentally. It is shown that the time complexity of the algorithms is independent of the number of control points, i.e. the compression ratio. Also a new morphological image decompression scheme is presented, the 8 kernels' method. Several decompressed images are presented. The best results are obtained using the Delaunay triangulation. The obtained image quality equals that of the DCT images with a 4 x 4

Image Segmentation Using Information Bottleneck Method

In image processing, segmentation algorithms constitute one of the main focuses of research. In this paper, new image segmentation algorithms based on a hard version of the information bottleneck method are presented. The objective of this method is to extract a compact representation of a variable, considered the input, with minimal loss of mutual information with respect to another variable, considered the output. First, we introduce a split-and-merge algorithm based on the definition of an information channel between a set of regions (input) of the image and the intensity histogram bins (output). From this channel, the maximization of the mutual information gain is used to optimize the image partitioning. Then, the merging process of the regions obtained in the previous phase is carried out by minimizing the loss of mutual information. From the inversion of the above channel, we also present a new histogram clustering algorithm based on the minimization of the mutual information loss, where now the input variable represents the histogram bins and the output is given by the set of regions obtained from the above split-and-merge algorithm. Finally, we introduce two new clustering algorithms which show how the information bottleneck method can be applied to the registration channel obtained when two multimodal images are correctly aligned. Different experiments on 2-D and 3-D images show the behavior of the proposed algorithms