Electronic Letters on Computer Vision and Image Analysis (ELCVIA - Universitat Autònoma de Barcelona)
Not a member yet
343 research outputs found
Sort by
Registration and analysis for images couple : Application to mammograms
Get PDFIn this thesis, the problem addressed is the development of a computer-aided diagnosis system (CAD) based on conjoint analysis of several images, and therefore on the comparison of these medical images. The particularity of our approach is to look for evolutions or aberrant new tissues in a given set, rather than attempting to characterize, with a strong a priori, the type of tissues. This problem allows to apprehend one aspect of the analysis of a medical file performed by experts which is the study of a case through comparison and evolution detection. The methodology proposed is carried out within the application context of the development of a CAD applied to mammograms. The first step when a couple of images are involved is to perform an adapted registration. Any automated comparison of signals requires an alignment of similar components present on the pictures, that is to say a registration phase, so that they occupy the same space on the two images. As the registration is never perfect, we must take into account the level of uncertainty and develop a comparison method able to distinguish registration error and real small differences between comparable tissues. In many applications, the assessment of similarity used during the registration step is also used in the interpretation step that yields to prompt suspicious regions. In our case, registration is assumed to match the spatial coordinates of similar anatomical elements
Reduced egomotion estimation drift using omnidirectional views
Get PDFEstimation of camera motion from a given image sequence is a common task for multi-view 3D computer vision applications. Salient features (lines, corners etc.) in the images are used to estimate the motion of the camera, also called egomotion. This estimation suffers from an error built-up as the length of the image sequence increases and this causes a drift in the estimated position. In this letter, this phenomenon is demonstrated and an approach to improve the estimation accuracy is proposed. The main idea of the proposed method is using an omnidirectional camera (360° horizontal field of view) in addition to a conventional (perspective) camera. Taking advantage of the correspondences between the omnidirectional and perspective images, the accuracy of camera position estimates can be improved. In our work, we adopt the sequential structure-from-motion approach which starts with estimating the motion between first two views and more views are added one by one. We automatically match points between omnidirectional and perspective views. Point correspondences are used for the estimation of epipolar geometry, followed by the reconstruction of 3D points with iterative linear triangulation. In addition, we calibrate our cameras using sphere camera model which covers both omnidirectional and perspective cameras. This enables us to treat the cameras in the same way at any step of structure-from-motion. We performed simulated and real image experiments to compare the estimation accuracy when only perspective views are used and when an omnidirectional view is added. Results show that the proposed idea of adding omnidirectional views reduces the drift in egomotion estimation
How to separate between Machine-Printed/Handwritten and Arabic/Latin Words?
Get PDFThis paper gathers some contributions to script and its nature identification. Different sets of featureshave been employed successfully for discriminating between handwritten and machine-printed Arabic and Latin scripts. They include some well established features, previously used in the literature, and new structural features which are intrinsic to Arabic and Latin scripts. The performance of such features is studied towards this paper. We also compared the performance of three classifiers: Bayes (AODEsr), k-Nearest Neighbor (k-NN) and Decision Tree (J48) used to identify the script at word level. These classifiers have been chosen enough different to test the feature contributions. Experiments have been conducted with handwritten and machine-printed words, covering a wide range of fonts. Experimental results show the capability of the proposed features to capture differences between scripts and the effectiveness of the three classifiers. An average identification precision and recall rates of 98.72% was achieved, using a set of 58 features and AODEsr classifier, which is slightly better than those reported in similar works
Exploiting Multimedia Content: A Machine Learning Based Aproach
Get PDFThis thesis explores use of machine learning for multimedia content management involving single/multiple features, modalities and concepts. We introduce shape based feature for binary patterns and apply it for recognition and retrieval application in single and multiple feature based architecture. The multiple feature based recognition and retrieval frameworks are based on the theory of multiple kernel learning (MKL). A binary pattern recognition framework is presented by combining the binary MKL classifiers using a decision directed acyclic graph. The evaluation is shown for Indian script character recognition, and MPEG7 shape symbol recognition. A word image based document indexing framework is presented using the distance based hashing (DBH) defined on learned pivot centres. We use a new multi-kernel learning scheme using a Genetic Algorithm for developing a kernel DBH based document image retrieval system. The experimental evaluation is presented on document collections of Devanagari, Bengali and English scripts. Next, methods for document retrieval using multi-modal information fusion are presented. Text/Graphics segmentation framework is presented for documents having a complex layout. We present a novel multi-modal document retrieval framework using the segmented regions. The approach is evaluated on English magazine pages. A document script identification framework is presented using decision level aggregation of page, paragraph and word level prediction. Latent Dirichlet Allocation based topic modelling with modified edit distance is introduced for the retrieval of documents having recognition inaccuracies. A multi-modal indexing framework for such documents is presented by a learning based combination of text and image based properties. Experimental results are shown on Devanagari script documents. Finally, we have investigated concept based approaches for multimedia analysis. A multi-modal document retrieval framework is presented by combining the generative and discriminative modelling for exploiting the cross-modal correlation between modalities. The combination is also explored for semantic concept recognition using multi-modal components of the same document, and different documents over a collection. An experimental evaluation of the framework is shown for semantic event detection in sport videos, and semantic labelling of components of multi-modal document images
Interactive and audience-adaptive information interfaces
Get PDFIn the doctoral thesis we developed an interactive and user-adaptive information interface based on computer vision and machine learning methods. By using a camera-enhanced digital signage display we employed real-time computer vision algorithms to extract temporal, spatial, and demographic features of the observers, which are further used for observer specific broadcasting of digital signage contents. The algorithms were chosen and modified to optimize the balance between accuracy and time complexity, subjected to design-aim to perform in real-time and using conventional hardware. More particularly, we used the Mixture of Gaussians method for background segmentation, Viola & Jones method for face detection algorithm, Active Appearance Models for face alignment and POSIT algorithm for head pose estimation. The developed interface is used as the key research tool to explore three currently open problems in the field of human-computer interaction: dynamic anamorphosis, quantitative audience measurement study of digital signage in real-world environment, and modeling of the purchase decision process. In the first study, we developed a new interactive computer vision based method which adapts image projection to the changing position of the observer so that wherever the observer moves, he sees the same undeformed image. We call this capacity dynamic anamorphosis. We formalized the anamorphic transformation and proposed a real-time algorithm for tracking the 3D position of the observer\u27s eyes and the re-computation of the anamorphic deformation. As an interesting application, we show that dynamic anamorphosis could be used to improve eye-contact in videoconferencing. In the second study, we used the developed interface to perform a quantitative audience measurement field study, which evaluates user attention. Temporal metrics of a person\u27s dwell time, display in-view time and attention time are extracted using real-time image analysis. The system also determines demographic metrics of the gender and age group based on images of faces. The digital signage display was deployed in a real-world environment of a clothing boutique, where demographic and viewership data of 1294 store customers were recorded, manually verified and analysed. The analysis shows that 35% of customers specifically looked-at the display, having the average attention time of 0.7 s. Interestingly, the attention time was substantially higher for men (1.2 s) than for women (0.4 s).In the third study, the interface is applied to model the purchase decision process, which is an interdisciplinary study, where data collected with the developed interface and subjected to machine learning are combined to model and analyze the decision and roles in a purchasing process. Finally, more generally, the developed system presents a contribution to the field of human-computer interaction and shows further possibilities for scientific use and applications, such as open problem of display blindness, development of new interactive methods for broadcasting of relevant content, and quantitative analysis of user behavior
Algorithms for selecting parameters of combination of acyclic adjacency graphs in the problem of texture image processing
Get PDFNowadays the great interest of researchers in the problem of processing the interrelated data arrays including images is retained. In the modern theory of machine learning, the problem of image processing is often viewed as a problem in the field of graph models. Image pixels constitute a unique array of interrelated elements. The interrelations between array elements are represented by an adjacency graph. The problem of image processing is often solved by minimizing Gibbs energy associated with corresponding adjacency graphs. The crucial disadvantage of Gibbs approach is that it requires empirical specifying of appropriate energy functions on cliques. In the present work, we investigate a simpler, but not less effective model, which is an expansion of the Markov chain theory. Our approach to image processing is based on the idea of replacing the arbitrary adjacency graphs by tree-like (acyclic in general) ones and linearly combining of acyclic Markov models in order to get the best quality of restoration of hidden classes. In this work, we propose algorithms for tuning combination of acyclic adjacency graphs
3D Scene Modeling And Understanding From Image Sequences
Get PDFA new method for 3D modeling is proposed, which generates a content-based 3D mosaic (CB3M) representation for long video sequences of 3D, dynamic urban scenes captured by a camera on a mobile platform. In the first phase, a set of parallel-perspective (pushbroom) mosaics with varying viewing directions is generated to capture both the 3D and dynamic aspects of the scene under the camera coverage. In the second phase, a unified patch-based stereo matching algorithm is applied to extract parametric representations of the color, structure and motion of the dynamic and/or 3D objects in urban scenes, where a lot of planar surfaces exist. Multiple pairs of stereo mosaics are used for facilitating reliable stereo matching, occlusion handling, accurate 3D reconstruction and robust moving target detection. The outcome of this phase is a CB3M representation, which is a highly compressed visual representation for a dynamic 3D scene, and has object contents of both 3D and motion information. In the third phase, a multi-layer based scene understanding algorithm is proposed, resulting in a planar surface model for higher-level object representations. Experimental results are given for both simulated and several different real video sequences of large-scale 3D scenes to show the accuracy and effectiveness of the representation. We also show the patch-based stereo matching algorithm and the CB3M representation can be generalized to 3D modeling with perspective views using either a single camera or a stereovision head on a ground mobile platform or a pedestrian. Applications of the proposed method include airborne or ground video surveillance, 3D urban scene modeling, traffic survey, transportation planning and the visual aid for perception and navigation of blind people
Optical Flow in Driver Assistance Systems
Get PDFIt is an extended abstract of the PhD thesis titled "Optical Flow in Driver Assistance Systems"
Polyp Localization and Segmentation in Colonoscopy Images by Means of a Model of Appearance for Polyps
Get PDFColorectal cancer is the fourth most common cause of cancer death worldwide and its survival rate depends on the stage in which it is detected on hence the necessity for an early colon screening. There are several screening techniques but colonoscopy is still nowadays the gold standard, although it has some drawbacks such as the miss rate. Our contribution, in the field of intelligent systems for colonoscopy, aims at providing a polyp localization and a polyp segmentation system based on a model of appearance for polyps. To develop both methods we define a model of appearance for polyps, which describes a polyp as enclosed by intensity valleys. The novelty of our contribution resides on the fact that we include in our model aspects of the image formation and we also consider the presence of other elements from the endoluminal scene such as specular highlights and blood vessels, which have an impact on the performance of our methods. In order to develop our polyp localization method we accumulate valley information in order to generate energy maps, which are also used to guide the polyp segmentation. Our methods achieve promising results in polyp localization and segmentation. As we want to explore the usability of our methods we present a comparative analysis between physicians fixations obtained via an eye tracking device and our polyp localization method. The results show that our method is indistinguishable to novice physicians although it is far from expert physicians
Automatic building detection and land use classification in urban areas using multispectral high-spatial resolution imagery and LiDAR data
Get PDFUrban areas areimportant environments, accounting for approximately half the population of theworld. Cities attract residents partly because they offer ample opportunitiesfor development, which often results in urban sprawl and its complex environmentalimplications. It is therefore necessary to develop technologies andmethodologies that permit monitoring the effects of various problems that havebeen or are thought to be associated with urban sprawl. These technologieswould facilitate the adoption of policies seeking to minimize the negativeeffects of urban sprawl. Solutions require a precise knowledge of the urbanenvironment under consideration to enable the development of more efficienturban zoning plans. The high dynamism of urban areas produces seeminglycontinuous alterations of land cover and use; consequently, cartographicinformation becomes quickly and is oftentimes outdated. Hence, the availabilityof detailed and up-to-date cartographic and geographic information is imperativefor an adequate management and planning of urban areas. Usually the process ofcreating land-use/land-cover maps of urban areas involves field visits andclassical photo-interpretation techniques employing aerial imagery. Thesemethodologies are expensive, time consuming, and also subjective. Digital imageprocessing techniques help reduce the volume of information that needs to bemanually interpreted.The aim of thisstudy is to establish a methodology to automatically detect buildings and toautomatically classify land use in urban environments using multispectralhigh-spatial resolution imagery and LiDAR data. These data were acquired in theframework of the Spanish National Plan for Airborne Orthophotographs, having beenavailable for public Spanish administrations.Two mainapproaches for automatic building detection and localization using high spatialresolution imagery and LiDAR data are evaluated The thresholding-based approachis founded on the establishment of two threshold values: one is the minimumheight to be considered as a building, defined using the LiDAR data; the other isthe presence of vegetation, defined with the spectral response. The otherapproach follows the standard scheme of object-based image classification:segmentation, feature extraction and selection, and classification, hereperformed using decision trees. In addition, the effect of including contextualrelations with shadows in the building detection process is evaluated. Qualityassessment is performed at both area and object levels. Area-level assessments evaluatethe building delineation performance whereas object-level assessments evaluatethe accuracy in the spatial location of individual buildings.Urban land-useclassification is achieved by applying object-based image analysis techniques.Objects are defined using the boundaries of cadastral plots. The plots were characterizedto achieve the classification by employing a descriptive feature setspecifically designed to describe urban environments. The proposed descriptivefeatures aim to emulate human cognition by numerically quantifying theproperties of the image elements and so enable each to be distinguishable.These features describe each plot as a single entity based on several aspectsthat reflect the information used: spectral, three-dimensional, and geometrictypologies. In addition, a set of contextual features at both the internal andexternal levels is defined. Internal context features describe an object withrespect to the land cover types contained within the plots, which were, in thiscase, buildings and vegetation. External context features characterise eachobject by considering the common properties of adjacent objects that, whencombined, create an aggregate in a higher level than plot level: urban blocks.Results show that thresholding-based building detection approachperforms better in the different scenarios assessed. This method produces amore accurate building delineation and object detection than the object-basedclassification method. The building type appears as a key factor in thebuilding detection performance. Thus, urban and industrial areas show betteraccuracies in detection metrics than suburban areas, due to the small size ofsuburban constructions, combined with the prominent presence of trees insuburban classes, hindering the building detection process. The relationsbetween buildings and shadows improve the object-level detection, removingsmall objects erroneously detected as buildings that negatively affect to thequality indices.Classificationtest results show that internal and external context features complement theimage-derived features, improving the classification accuracy values of urbanclasses, especially between classes that show similarities in their image-basedand three-dimensional features. Context features enable a superiordiscrimination of suburban building typologies, of planned urban areas andhistorical areas, and also of planned urban areas and isolated buildings.The outcomes showthat these automatic methodologies are especially suitable for computing usefulinformation for constructing and updating land-use/land-cover geospatialdatabases. Digital image processing-based methodologies provide better resultsthan visual interpretation-based methods. Thus, automatic building detectiontechniques produce a superior estimation of built-up surface in an objectivemanner, independent of human operators. The combination of building detectionand automatic classification of land use in urban areas enable the distinguishingand describing of different urban typologies, contributing to greater accuracyand information than standard visual interpretation-based techniques. Theproposed methodology, based on an automated descriptive feature extraction fromLiDAR images and data, is appropriate for city mapping, urban landscapecharacterisation and management, and the updating of geospatial databases, allof which provide novel tools to increase the frequency and efficiency of thestudy of complex urban areas