BieColl - Bielefeld eCollections
Not a member yet
1004 research outputs found
Sort by
First steps towards an intentional vision system
Get PDFContrary to many standard vision systems which proceed in a cascaded feedforward manner, imposing a fixed order in the sequence of visual operations like detection preceding segmentation and classification, we develop here the idea of a vision system that flexibly controls the order and accessibility of visual processes during operation. Vision is hereby understood as the dynamic process of adaptation of visual parameters and modules as a function of underlying goals or intentions. This perspective requires a specific architectural organization, since vision is then a continuous balance between the sensory stimulation and internally generated information. In this paper we present the concept and the necessary main ingredients and show first steps towards the implementation of a real-time intentional vision system
Machine Perception using a Blackboard Architecture
Get PDFHere we present ongoing research in the application of symbolic reasoning to perception in general and vision in particular. Perception is treated as the combination of the possibly contradictory outputs of many specialized processes which communicate via a blackboard data structure. It is demonstrated that our design allows for bottom-up, horizontal and top-down information flow. Significant progress towards the analysis of unstructured scenes has been made. The principles involved have been explored experimentally and preliminary results are presented
Online Learning of Objects and Faces in an Integrated Biologically Motivated Architecture
Get PDFWe present a biologically motivated integrated vision system that is capable of online learning of several objects and faces in a unified representation. The training is unconstrained in the sense that arbitrary objects can be freely presented in front of a stereo camera system and labeled by speech input. We combine biological principles such as appearance-based representation in topographical feature detection hierarchies and context-driven transfer between different levels of object memory. The learning is driven by interactively sharing attention between user and system. It is fully online and avoids an artificial separation of the interaction into training and test phases
Detection of Anomalies and Novelties in Time Series with Self-Organizing Networks
Get PDFThis paper introduces the DANTE project: Detection of Anomalies and Novelties in Time sEries with self-organizing networks. The goal of this project is to evaluate several self-organizing networks in the detection of anomalies/novelties in dynamic data patterns. For this purpose, we first describe three standard clustering-based approaches which uses well-known self-organizing neural architectures, such as the SOM and the Fuzzy ART algorithms, and then present a novel approach based on the Operator Map (OPM) network. The OPM is a generalization of the SOM where neurons are regarded as temporal filters for dynamic patters. The OPM is used to build local adaptive filters for a given nonstationary time series. Non-parametric confidence intervals are then computed for the residuals of the local models and used as decision thresholds for detecting novelties/anomalies. Computer simulations are carried out to compare the performances of the aforementioned algorithms
Dimensionality Reduction of very large document collections by Semantic Mapping
Get PDFThis paper describes improving in Semantic Mapping, a feature extraction method useful to dimensionality reduction of vectors representing documents of large text collections. This method may be viewed as a specialization of the Random Mapping, method proposed in WEBSOM project. Semantic Mapping, Random Mapping and Principal Component Analysis (PCA) are applied to categorization of document collections using Self-Organizing Maps (SOM). Semantic Mapping generated document representation as good as PCA and much better than Random Mapping
The activation frequency self-organizing map
Get PDFIn the self-organizing map (SOM), the best matching units (BMUs) affect neurons as a function of distance and the learning parameter. Here we study the effects in SOM when a new parameter in the learning rule, the activation frequency, is included. This parameter is based on the relative frequency by which each neuron is included in each BMU's neighborhood, so there is an individual memory (synapse strength) of the activation received from each neuron. The parameter leads to non-radial influence areas for BMUs, what is a more realistic feature observed in the brain cortex which modifies the map formation dynamics, including the fact that the weight vector for BMU may not be the closest one to the input stimulus after weight adaptation. Also, two error measures are lower for the maps trained with this model than those obtained with SOM, as shown in several experiments with six data sets
A Biomimetic Vision Architecture
Get PDFThe goal of biomimetic vision is to build artificial vision systems that are analogous to the human visual system. This paper presents a software architecture for biomimetic vision in which every major component is clearly defined in terms of its function and interface, and where every component has a analog in the regional functional anatomy of the human brain. We also present an end-to-end vision system implemented within this framework that learns to recognize objects without human supervision
A Constrained Alternating Optimization Framework for Feature Matching
Get PDFThis paper proposes a constrained alternating optimization framework to tackle the feature matching problem with partial matching and multiple matching. We model the difference between pairing features as the result of a transformation followed an uncertainty distribution. Based on this modeling, transformation estimation and feature matching are performed alternately from initial matching: the transformation is updated according to the matching, and the matching is updated according to the transformation and the uncertainty distribution. A pruning operation is further presented to reduce the search space of initial matching. In the proposed framework, we develop a B-spline curve feature matching algorithm for hand-gesture based text input, and a line feature matching algorithm which is tested for three applications: model-based recognition, image registration, and stereo matching. The experimental results for two algorithms are reported
A New Method and Toolbox for Easily Calibrating Omnidirectional Cameras
Get PDFIn this paper, we focus on calibration of central omnidirectional cameras, both dioptric and catadioptric. We describe our novel camera model and algorithm and provide a practical Matlab Toolbox, which implements the proposed method. Our method relies on the use of a planar grid that is shown by the user at different unknown positions and orientations. The user is only asked to click on the corner points of the images of this grid. Then, calibration is quickly and automatically performed. In contrast with previous approaches, we do not use any specific model of the omnidirectional sensor. Conversely, we assume that the imaging function can be described by a polynomial approximation whose coefficients are estimated by solving a linear least squares minimization problem followed by a non-linear refinement. The performance of the approach is shown through several calibration experiments on both simulated and real data. The proposed algorithm is implemented as a Matlab Toolbox, which allows any inexpert user to easily calibrate his own camera. The toolbox is completely Open Source and is freely downloadable from the author's Web page
A Three-Level Computational Attention Model
Get PDFThis article deals with a biologically-motivated three-level computational attention model architecture based on the rarity and the information theory framework. It mainly focuses on a low-level step which aims in fastly highlighting important areas and a middle-level step which analyses the behaviour of the detected areas. Their application on both still images and videos provide results to be used by the third high-level step