This master thesis is focused on parametrization of image point neighborhood. Some methods for point localization and point descriptors are described and summarized. Gabor filter is described in detail. The practical part of thesis is chiefly concerned with particle filter tracking system. The weight of each particle is determined by the Gabor filter.

Vision based object recognition and localization have been studied widely in recent years. Often the initial step in such tasks is detection of interest points from a grey-level image. The current state-of-the-art algorithms in this domain, like Scale Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF) suffer from low execution speeds on a GPU(graphic processing unit) based system. Generally the performance of these algorithms on a GPU is below real-time due to high computational complexity and data intensive nature and results in elevated power consumption. Since real-time performance is desirable in many vision based applications, hardware based feature detection is an emerging solution that exploits inherent parallelism in such algorithms to achieve significant speed gains. The efficient utilization of resources still remains a challenge that directly effects the cost of hardware. This work proposes a novel hardware architecture for scale-space extrema detection part of the SIFT algorithm. The implementation of proposed architecture for Xilinx Virtex-4 FPGA and its evaluation are also presented. The implementation is sufficiently generic and can be adapted to different design parameters efficiently according to the requirements of application. The achieved system performance exceeds real-time requirements (30 frames per second) on a 640 x 480 image. Synthesis results show efficient resource utilization when compared with the existing known implementations.

Relatively complex Augmented Reality (AR) algorithms are becoming widely available due to advancements in affordable mobile computer hardware. To take advantage of this a new method is developed for tracking 2D regions without a prior knowledge of an environment and without developing a computationally expensive world model. In the method of this paper, affinely invariant planar regions in a scene are found using the Maximally Stable Extremal Region (MSER) detector. A region is selected by the user to define a search space, and then the Scale Invariant Feature Transform (SIFT) is used to detect affine invariant keypoints in the region. If three or more keypoint matches across frames are found, the affine transform A of a region is calculated. A 2D image is then transformed by A causing it to appear stationary on the 2D region being tracked. The search region is tracked by transforming the previous search region by A, defining a new location, size, and shape for the search region. Testing reveals that the method is robust to tracking planar surfaces despite affine changes in the geometry of a scene. Many real world surfaces provide adequate texture for successful augmentation of a scene. Regions found multiple frames are consistent with one another, with a mean cross-correlation of 0.608 relating augmented regions. The system can handle up to a 45° out of plane viewpoint change with respect to the camera. Although rotational changes appear to skew the affine transform slightly, translational and scale based have little distortion and provide convincing augmentations of graphics onto the real world.

This is a study exploring how robust one feature descriptors, scale invariant feature transform (SIFT), and one feature detector, feature accelerated segmentation test (FAST), are in terms of handling fetoscopic acquired data when mosaicing. Today’s treatment of severe Twin-to-Twin Transfusion Syndrome at Karolinska University Hospital is fetoscopic guided laser occlusion of chorioangiopagous vessels (FLOC) where intersecting blood vessels causing a transfusion (Anastomoses) in between the fetuses are occluded. These blood vessels are located somewhere on the placenta. The fetoscopy includes navigation of a relatively large area where the field of view (FOV) is limited. The limited FOV during the fetoscopy makes it cumbersome to navigate and identify intersected blood vessels. The motivation of this study is to explore ways of dealing with the complications during FLOC by mosaicing an overview of the placenta that can be used as an assisting map to make the procedure safer by improving navigation of the fetoscope and identification of blood vessels during FLOC. In this study, the steps of mosaicing are defined based on mosaicing frameworks to explore how these methods perform in terms of being able to mosaic a map of the placenta. The methods have been tested on non-fetoscopic acquired data as well as fetoscopic acquired data to create a relative measure in between the two. Three tests on non- fetoscopic data were performed to explore how well the methods handled mosaicing of data with distinctive characteristics. The same methods were then tested on unprocessed fetoscopic data before being tested on preprocessed fetoscopic data to see if the results were affected by external preprocessing. The results showed that there were differences in between the methods. SIFT and FAST showed that they have potential of mosaicing non-fetoscopic data of varying extent. SIFT gave an impression of being more robust during all of the tests. SIFT especially performed better during the tests on data with few potential keypoints which is an advantage when speaking of fetoscopic acquired data. SIFT also managed to mosaic a larger area than FAST when mosaicing preprocessed fetoscopic data. Preprocessing the data improved the mosaicing when using SIFT but further improvements are needed. / Denna studie utforskar hur robust en intressepunktsbeskrivare, scale invariant feature transform (SIFT), och enintressepunktsdetektor, feature accelerated segmentation test (FAST), hanterar digitala bilder insamlade av ett fetoskop med syfte att sy ihop dessa till en översiktskarta. Dagens behandling av tvillingtransfusionsyndrom vid Karolinska Universitetssjukhuset är fetoscopic guided laser occlusion of chorioangiopagous vessels (FLOC). Under denna fetoskopi bränner man och därmed blockerar korsande blodkärl som orsakar en transfusion (anastomoses) och obalans i blodomloppet mellan två tvillingfoster. Dessa blodkärl är lokaliserade på placentan. Fetoskopin omfattar navigering av en relativt stor area med ett begränsat synfält. Det begränsade synfältet under FLOC gör det svårt att orientera fetoskopet och identifiera korsande blodkärl som orsakar transfusionen. Syftet med studien är att utforska ett sätt att hantera komplikationerna med FLOC igenom att utforska sätt att skapa en översiktskarta av placentan under FLOC. Denna översiktskarta kan nyttjas under FLOC och därmed göra proceduren säkrare igenom att kartan underlättar orienteringen av fetoskopet och identifiering av orsakande blodkärl. I denna studie är stegen för att skapa en översiktskarta baserade på olika datorseende ramverk för att se hur dessa tillvägagångssätt presterar när det gäller att skapa en översiktskarta. Metoderna för att skapa en översiktskarta har testats på data insamlad med webkamera och data insamlad med fetoskop för att skapa en relativ uppfattning om hur de står sig beroende på indata. Tre tester på data insamlad med webkamera genomfördes för att utforska hur väl metoderna hanterade data med många potentiella intressepunkter, rörelse orsakad av handhållen enhet/kamera, repetitiva mönster, översiktskartor som resulterande i större upplösning, och liten möjlighet att hitta intressepunkter. Samma metoder testades sedan på icke behandlad data insamlad med fetoskop innan den testades på förbehandlad data insamlad med fetoskop för att se förbehandlingensnödvändighet och prestation. Resultaten visar att det är skillnader mellan de två metoderna använda i denna studie, både när det gäller data insamlad med fetoskop och webkamera. SIFT och FAST visar potential av olika grad när det gäller att skapa en översiktskarta med data insamlad av webkamera. SIFT visade sig vara mer robust under alla tester inklusive data insamlad med fetoskop. SIFT presterade speciellt bättre under testen som omfattade få antal möjliga intressepunkter vilket är en fördel när det gäller data insamlad med fetoskop. SIFT lyckades också skapa översiktskartor med större area än FAST när förbehandlad fetoskopisk data testades. När det gäller SIFT så visade resultaten en förbättring när data insamlad med fetoskop förbehandlades men att ytterligare förbättringar är nödvändiga.

This study investigates the performance of two keypoint detection algorithms, SIFTand LoFTR, for vehicle re-recognition on a 2+1 road in Täby, utilizing three differentmethods: proportion of matches, ”gates” based on the values of the features andSupport Vector Machines (SVM). Data was collected from four strategically placedcameras, with a subset of the data manually annotated and divided into training,validation, and testing sets to minimize overfitting and ensure generalization. TheF1-score was used as the primary metric to evaluate the performance of the variousmethods. Results indicate that LoFTR outperforms SIFT across all methods, with theSVM method demonstrating the best performance and adaptability. The findings havepractical implications in security, traffic management, and intelligent transportationsystems, and suggest directions for future research in real-time implementation andgeneralization across varied camera placements.

No description available.

No description available.

One particular challenge of brain connectomics deals with inferring differences in the brain due to diseases such as Alzheimer's. More specifically, structural connectomics aims at investigating the connectivity between regions in the brain based on the distribution of neuronal fibers. The first step in generating structural connectomes is to perform tractography reconstruction on diffusion MRI (dMRI) data, to extract the most likely pathways of neural fibers. However, current tractography reconstruction algorithms suffer from having high sensitivity and low specificity. Thus, the following steps  of creating, analyzing and deriving graphs metrics from connectivity maps based on tractography impair the reliable assessment of structural connectivity. A promising method to improve tractography and subsequent structural connectomes is to apply tractogram filtering methods. In this study, the impact of tractogram filtering on structural connectomics and derived graph measures of subjects with mild cognitive impairment (MCI), specifically using spherical-deconvolution informed filtering of tractograms (SIFT), is experimentally examined. Moreover, the study also aims at inferring the effects of tractogram filtering in machine-learning based classification of the aforementioned structural connectomes. The pipeline in this experimental setup uses registration tools from FSL, tractography tools from MRTrix3Tissue as well as Keras for classification. The results from the given experiments show, that graph measures such as nodestrength and betweenness centrality are altered for the individual nodes. This leads to new connectomes with nodes, which are more important after tractogram filtering. This effect was also seen in connectomes weighted by fractional anisotropy (FA), mean diffusivity (MD) and radial diffusivity (RD). Moreover, structural connectomes based on filtered tractograms yield a higher classification performance. The best classification performance was reached with 88.65% on raw connectomes. Limiting factors in this experimental setup are identified as the small number of subjects at hand and computation time and the errors introduced by image registration and tractography parameterization.

Avec le développement du numérique, le nombre d'images stockées dans les bases de données a beaucoup augmenté. L'indexation des images et la recherche d'information dans les bases d'images sont plus compliquées que dans le cas de documents textuels Des méthodes d'indexation déjà utilisées en analyse de données textuelles sont proposées pour traiter des images. Pour transférer les résultats de l'analyse de données textuelles aux images, il est nécessaire d'utiliser de nouvelles caractéristiques : les mots visuels et on considère les images comme documents. Nous nous intéressons au problème d'indexation et de recherche d'information dans des grandes bases de données d'images à l'aide de méthodes d'analyse de données comme l'Analyse Factorielle des Correspondances (AFC). Nous proposons d'abord une utilisation astucieuse des indicateurs de l'AFC pour accélérer la recherche après l'avoir adaptée aux images. Nous nous intéressons ensuite au problème du passage à l'échelle de l'AFC. Pour ce faire, nous proposons un algorithme d'AFC incrémentale pour traiter de grands tableaux de données et la parallélisation de cet algorithme sur processeurs graphiques (GPU). Nous développons aussi une version parallèle de notre algorithme de recherche qui utilise les indicateurs de l'AFC sur GPU. Puis nous associons l'AFC à d'autres méthodes comme la Mesure de Dissimilarité Contextuelle ou les forêts aléatoires pour améliorer la qualité de la recherche. Enfin, nous présentons un environnement de visualisation, CAViz, pour accompagner les traitements précédents.

Les systèmes de recherche d'images par le contenu (Content-Based Image Retrieval System - CBIRS) sont maintenant couramment utilisés comme mécanismes de filtrage contre le piratage des contenus multimédias. Ces systèmes utilisent souvent le schéma de description d'images SIFT pour sa bonne robustesse face à un large spectre de distorsions visuelles. Mais aucun de ces systèmes n'a encore abordé le problème du piratage à partir d'un point de vue ''sécurité''. Cette thèse a comme objectif d'analyser les CBIRS de ce point de vue sécurité. Il s'agit de comprendre si un pirate peut produire des distorsions visuelles perturbant les capacités de reconnaissances d'un système en créant ces distorsions en fonctions des techniques que ce système utilise. Tout d'abord, nous présentons les failles de sécurité des composantes typiques d'un CBIRS : composantes description d'image, indexation et filtrage des faux positifs. Ensuite, nous présentons des attaques ciblant le schéma de description SIFT. Les attaques sont effectuées durant l'étape de détection de points d'intérêt et de calculs des descripteurs. Nous présentons également une attaque ciblant la mise en correspondance des images sur un critère de cohérence géométrique. Les expériences menées avec 100 000 images réelles confirment l'efficacité des attaques proposées.