Electronic Letters on Computer Vision and Image Analysis (ELCVIA - Universitat Autònoma de Barcelona)
Not a member yet
343 research outputs found
Sort by
On a Distributed Video Surveillance System to Track Persons in Camera Networks
Full text linkIn automated video surveillance applications, presenting the useful information to the human operators is a challenging task. Current systems usually require a prohibitive amount of human resources and lead to a quick decrease of the attention of the human operators through time, thus preventing them to catch the relevant events that may be worth to further investigate. In addition, when monitoring a wide area, it becomes hard to deploy a network of video sensors such that there are enough overlapping FoVs to cover every point of the environment. This leads to the development of video surveillance systems (VSS) that provide partial area coverage. As a result, “blind-gaps” between camera FoVs are introduced. One of the most interesting problems which such “blind-gaps” bring in is to re-identify the persons moving across disjoint FoVs.The contribution of the thesis is two-fold. First, an advanced VSS is designed to display the proper taskdependent information to operators that are monitoring a wide area. In particular, the system helps operators in the task of tracking persons across camera views. This raised the need for a system capable of re-identify the subjects moving through disjoint FoVs. This leads to the second contribution of the thesis, a distributed approach to address the challenges of the person re-identification problem
Learning of invariant object recognition in hierarchical neural networks using temporal continuity
Full text linkA lot of progress in the field of invariant object recognition has been
made in recent years using so called deep neural networks with several
layers to be trained which can learn patterns of increasing complexity. This
architectural feature can alreay be found in older neural models as, e.g.,
the Neocognitron and HMAX but also newer ones as Convolutional Nets. Additionally
researchers emphasized the importance of temporal continuity in input data
and devised learning rules utilizing it (e.g. the trace rule by F\"oldiak
used by Rolls in VisNet). Finally Jeff Hawkins collected a lot of these ideas
concerning functioning of the neocortex in a coherent framework and proposed
three basic principles for neocortical computations (later implemented in HTM):Learning of temporal sequences for creating invariance to transformations contained in the training data.Learning in a hierarchical structure, in which lower level knowledge can be reused in higher level context and thereby makes memory usage efficient.Prediction of future signals for disambiguation of noisy input by feedback.In my thesis I developed two related systems: the \emph{Temporal Correlation Graph} (TCG)
and the \emph{Temporal Correlation Net} (TCN). Both make use of these principles and
implement them in an efficient manner. The main aim was to create systems that are trained
mostly unsupervised (both) and can be trained online, which is possible with TCN. Both achieve
very good performance on several standard datasets for object recognition
Document Image Binarization Using Retinex and Global Thresholding
Full text linkDocument images are usually degraded in the course of photocopying, faxing, printing, or scanning. Degradation problems seems negligible to human eyes but can be responsible for an abrupt decline in accuracy by the current generation of optical character recognition (OCR) systems. In this paper we present binarization method based on retinex theory followed by a global threshold. High quality results in terms of visual criteria and OCR performance is produced compared to the previous works
DROP: A Data Reduction and Organization Paradigm and its Application in Image Analysis
Full text linkIn this paper, we deal with the problem of the annotation process in image analysis. This problem refers to the trade-off, wherein the human knowledge is indispensable for the success of the process and human\u27s time and effort are precious resources. Can the human annotate a minimum number of images and the classifier label the remaining ones with high accuracy? Active learning techniques have been investigated to answer this question. However, these techniques very often ignore the need for interactive response times during the active learning process. They usually adopt a common paradigm which is impractical considering large datasets. We propose an effective and efficient Data Reduction and Organization Paradigm for image analysis. In our paradigm, the proposed active learning methods should be able to reduce and/or organize the large dataset such that sample selection does not require to reprocess it entirely at each learning iteration. Moreover, it can be interrupted as soon as a desired number of samples from the reduced and organized dataset is identified. These methods show an increasing progress, first with data reduction only, and then with subsequent organization of the reduced dataset. Experimental results have demonstrated the robustness of the proposed paradigm using datasets from distinct applications and baseline approaches
An Evaluation of Perceptual Classification led by Cognitive Models in Traffic Scenes
Full text linkThe objects extraction and recognition constitute the most important link in the image processing and understanding, and it cannot be achieved without a solid objects organization during the processing through the learning mechanisms. Most often, both the response time and the accuracy are undeniable criteria for applications in this field. Actually, a vision system need to take into consideration these criteria, either in the structural, the methodological or in the algorithmic aspect. Thus, we consider that the ontological study at the domain and task levels, in the vision systems, has become essential in order to provide a substantial assistance to the multitudes of applications in image processing. Concerning the domain knowledge, several patterns for structuring were proposed to improve the objects representation and organization, they often advocate the precision aspect on time and on effort devoted to the recognition. In practical terms, clustering methods only focus on the accuracy aspect within a category, without considering the recognition aspect [1]. Thus, we propose in this study a new procedure of object categorization, which uses, according to the expertise in the domain, a fit evaluation that is able to adjust the level of partitioning. As a result, this procedure will find a compromise between the accuracy on the categories and the reduction of the supplied effort in recognition.
Interpreting the Structure of Single Images by Learning from Examples
Full text linkAn important problem in computer vision is the interpretation of the content of a single image. In our work we investigated the challenging case of recovering the underlying 3D structure of a scene from a single image, by learning from trainig data. Toward this, we developed a plane detection algorithm, which is able to find planar surfaces in a single still image and estimate their orientation with respect to the camera. This comprises two parts: a plane recognition stage, to classify individual regions as being planar or not, and to estimate their orienation; followed by a Markov-random field based segmentation stage to find distinct planes in the image. We also demonstrated an application of this to visual odometry, where single-image plane detection allows structure-rich maps to be built quickly. (Please note that this abstract does not appear in the submitted article itself, since that is itself an extended thesis abstract! But the above describes the main points of our work as described in our submission.
Toward a perceptual object recognition system
Full text link[1] demonstrated that humans are easily able to recognize an object in less than 0.5 seconds. Unfortunately,object recognition remains one of the most challenging problems in computer vision. Many algorithms basedon local approaches have been proposed in recent decades. Local approaches can be divided in 4 phases:region selection, region appearance description, image representation and classification [2]. Although thesesystems have demonstrated excellent performance, some weaknesses remain. The first limitation is in the region selection phase. Many existing techniques extract a large number of points/regions of interest. For instance, dense grids contain tens of thousands of points per image while interest point detectors often extract thousands of points. Furthermore, some studies have demonstrated that these techniques were not designed to detect the most pertinent regions for object recognition. There is only a weak correlation between the distribution of extracted points and eye fixations [3]. The second limitation mentioned in the literature concerns the region appearance description phase. The techniques used in this phase typically describe image regions using high-dimensional vectors [4]. For example, SIFT, the most popular descriptor for object recognition, produces a 128-dimensional vector per region [5].The main objective of this thesis is to propose a pipeline for an object recognition algorithm based on human perception which addresses the object recognition system complexity: query run time and memory allocation. In this context, we propose a filter based on a visual attention system [6] to address the problems of extracting a large number of points of interest using existing region selection techniques. We chose to use bottom-up visual attention systems that encode attentional fixations in a topographic map, known as a saliency map. This map serves as basis for generating a mask to select salient points according to human interest, from the points extracted by a region selection technique [7]. Furthermore, we addressed the problem of high dimensionality of descriptors in region appearance phase. We proposed a new hybrid descriptor representing the spatial frequency of some perceptual features, extracted by a visual attention system (color, texture, intensity [8]. This descriptor consist of a concatenation of energy measures computed at the output of a filter bank [9], at each level of the multi-resolution pyramid of perceptual features. This descriptor has the advantage of being lower dimensional than traditional descriptors.The test of our filtering approach, using Perreira da Silva system [10] as a filter on VOC2005, demonstrated that we can maintain approximately the same performance of an object recognition system by selecting only 40% of extracted points (using Harris-Laplace [11] and Laplacian [12]), while having an important reduction in complexity (40% reduction in query run time). Furthermore, evaluating our descriptor with an object recognition system using Harris-Laplace and Laplacian interest point detectors on VOC2007 database showed a slight decrease in performance ( 5% reduction of average precision) compared to the original system based on the SIFT descriptor, but with a 50% reduction in complexity. In addition, we evaluated our descriptor using a visual attention system as the region selection technique on VOC2005. The experiment showed a slight decrease in performance (3% reduction in precision), but a drastically reduced complexity of the system (with 5% reduction in query run-time and 70% in complexity).In this thesis, we proposed two approaches to manage the problems of complexity in object recognitionsystem. In future, it would be interesting to address the problems of the last two phases in object system: image representation and classification, by introducing perceptually plausible concepts such as deep learning techniques
Image Processing for Art Investigation
Full text linkRecent advances in digital image acquisition methods and the wide range of imaging modalities currently available have triggered museums to digitize their painting collections. Not only is this crucial for archival or dissemination purposes but it also enabled the digital analysis of the painting through its digital image counterpart. It also set in motion a cross-disciplinary collaboration between image analysis specialists, mathematicians, statisticians and art historians that have the common goal to develop algorithms and build a digital toolbox in support of art scholarship. Computer processing of digital images of paintings has become a fast growing and challenging field of research during the last few years. Our contribution to this research domain consists of a set of tools that are based on dimensionality reduction methods, sparse representations and dictionary learning techniques. These tools are used to assist in art related matters such as restoration, conservation, art history, material and structure characterization, authentication, dating and even style analysis. Since paintings are complex structures the analysis of all pictorial layers and the support requires a multimodal set of high-resolution image acquisitions. The presented research can broadly be subdivided into three main fields. The first one is the digital enhancement of painting acquisitions in order to assist the art specialist in his professional assessment of the painting. The second main field of research is the automated detection of cracks within the Ghent Altarpiece, which is meant to help in the delicate matter of the conservation of this exceptional masterpiece but also as guidance during its current campaign of restoration. The last field consists of a set of methods that can be deployed in art forensics. These methods consist of the characterization of canvas, the analysis of multispectral imagery of a painting and even the objective quantification of the style of a particular artist.
Image Analysis and Processing with Applications in Proteomics and Medicine
Full text linkThis thesis introduces unsupervised image analysis algorithms for the segmentation of several types of images, with an emphasis on proteomics and medical images. Segmentation is a challenging task in computer vision with essential applications in biomedical engineering, remote sensing, robotics and automation. Typically, the target region is separated from the rest of image regions utilizing defining features including intensity, texture, color or motion cues. In this light, multiple segments are generated and the selection of the most significant segments becomes a controversial decision as it highly hinges on heuristic considerations. Moreover, the separation of the target regions is impeded by several daunting factors such as: background clutter, the presence of noise and artifacts as well as occlusions on multiple target regions. This thesis focuses on image segmentation using deformable models and specifically region-based Active Contours (ACs) because of their strong mathematical foundation and their appealing properties
From pixels to gestures: learning visual representations for human analysis in color and depth data sequences
Full text linkThe visual analysis of humans from images is an important topic of interest due to its relevance to many computer vision applications likepedestrian detection, monitoring and surveillance, human-computer interaction, e-health or content-based image retrieval, among others.In this dissertation we are interested in learning different visual representations of the human body that are helpful for the visual analysis of humans in images and video sequences. To that end, we analyze both RGB and depth image modalities and address the problem from three different research lines, at different levels of abstraction; from pixels to gestures: human segmentation, human pose estimation and gesture recognition.First, we show how binary segmentation (object vs. background) of the human body in image sequences is helpful to remove all the background clutter present in the scene. The presented method, based on Graph cuts optimization, enforces spatio-temporal consistency of the produced segmentation masks among consecutive frames. Secondly, we present a framework for multi-label segmentation for obtaining much more detailed segmentation masks: instead of just obtaining a binary representation separating the human body from the background, finer segmentation masks can be obtained separating the different body parts.At a higher level of abstraction, we aim for a simpler yet descriptive representation of the human body. Human pose estimation methods usually rely on skeletal models of the human body, formed by segments (or rectangles) that represent the body limbs, appropriately connected following the kinematic constraints of the human body. In practice, such skeletal models must fulfill some constraints in order to allow for efficient inference, while actually limiting the expressiveness of the model. In order to cope with this, we introduce a top-down approach for predicting the position of the body parts in the model, using a mid-level part representation based on Poselets.Finally, we propose a framework for gesture recognition based on the bag of visual words framework. We leverage the benefits of RGB and depth image modalities by combining modality-specific visual vocabularies in a late fusion fashion. A new rotation-variant depth descriptor is presented, yielding better results than other state-of-the-art descriptors. Moreover, spatio-temporal pyramids are used to encode rough spatial and temporal structure. In addition, we present a probabilistic reformulation of Dynamic Time Warping for gesture segmentation in video sequences. A Gaussian-based probabilistic model of a gesture is learnt, implicitly encoding possible deformations in both spatial and time domains