Proposes a neural network based on differential Gabor filters for computing the image flow. The approach attempts to overcome the limitation of the spatio-temporal frequency models by taking time derivatives of the Ga...
详细信息
Proposes a neural network based on differential Gabor filters for computing the image flow. The approach attempts to overcome the limitation of the spatio-temporal frequency models by taking time derivatives of the Gabor responses as the carrier of visual motion information. A differential Gabor filter is a linear filter with the spatial derivative of a Gabo elemental function as its impulse response function. The authors derive a rigorous scheme for computing image motion. Based on this computational scheme, they present the architecture of a neural network system for visual motion. The computational model effectively bypasses the certainty constraint that severely limits the accuracy of the spatio-temporal frequency models, and avoids the time dimension integration required by the spatio-temporal frequency models. Experimental results show that the differential Gabor filter model performs better than the existing models.< >
Several texture segmentation algorithms based on deterministic and stochastic relaxation principles, and their implementation on parallel networks, are described. The segmentation process is posed as an optimization p...
详细信息
Several texture segmentation algorithms based on deterministic and stochastic relaxation principles, and their implementation on parallel networks, are described. The segmentation process is posed as an optimization problem and two different optimality criteria are considered. The first criterion involves maximizing the posterior distribution of the intensity field given the label field (maximum a posteriori estimate). The posterior distribution of the texture labels is derived by modeling the textures as Gauss Markov random fields (GMRFs) and characterizing the distribution of different texture labels by a discrete multilevel Markov model. A stochastic learning algorithm is proposed. This iterated hill-climbing algorithm combines fast convergence of deterministic relaxation with the sustained exploration of the stochastic algorithms, but is guaranteed to find only a local minimum. The second optimality criterion requires minimizing the expected percentage of misclassification per pixel by maximizing the posterior marginal distribution, and the maximum posterior marginal algorithm is used to obtain the corresponding solution. All these methods implemented on parallel networks can be easily extended for hierarchical segmentation; results of the various schemes in classifying some real textured images are presented.< >
The stochastic relaxation algorithm of S. Geman and D. Geman (IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-6, pp. 721-741, 1984), which is based on the Markov random-field (MRF) model of images, was implemented ...
详细信息
The stochastic relaxation algorithm of S. Geman and D. Geman (IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-6, pp. 721-741, 1984), which is based on the Markov random-field (MRF) model of images, was implemented on a hypercube parallel computer for contour extraction of human face images. The local energy parameters that define the MRF model were estimated by the learning algorithm recently proposed by the authors, while supervised by a desirable object contour as a teaching signal. The constrained optimization method recently developed by D. Geman (imagevision Comput., vol. 5, no. 2, pp. 61-65, 1987) was utilized to avoid flaws in a fast simulated annealing. The contours extracted by the learning stochastic relaxation method were systematically compared with those obtained by several conventional edge detection methods. The contours extracted by the authors' method include few discontinuity points and small amounts of noise, and faithfully represent the contours of the original image. The authors propose a new algorithm, multiple-level, multiple-resolution MRF, which is an extension of the original MRF. This model can incorporate a priori knowledge about the global structures in images and still be implemented in a local and parallel mode.
Cellular neural Networks (CNNs) constitute a class of nonlinear, recurrent and locally coupled arrays of identical dynamical cells that operate in parallel. ANALOG chips are being developed for use in applications ...
详细信息
ISBN:
(数字)9781475732207
ISBN:
(纸本)9780792378914;9781441949882
Cellular neural Networks (CNNs) constitute a class of nonlinear, recurrent and locally coupled arrays of identical dynamical cells that operate in parallel. ANALOG chips are being developed for use in applications where sophisticated signalprocessing at low power consumption is required.;signalprocessing via CNNs only becomes efficient if the network is implemented in analog hardware. In view of the physical limitations that analog implementations entail, robust operation of a CNN chip with respect to parameter variations has to be insured. By far not all mathematically possible CNN tasks can be carried out reliably on an analog chip; some of them are inherently too sensitive. This book defines a robustness measure to quantify the degree of robustness and proposes an exact and direct analytical design method for the synthesis of optimally robust network parameters. The method is based on a design centering technique which is generally applicable where linear constraints have to be satisfied in an optimum way.;processing speed is always crucial when discussing signal-processing devices. In the case of the CNN, it is shown that the setting time can be specified in closed analytical expressions, which permits, on the one hand, parameter optimization with respect to speed and, on the other hand, efficient numerical integration of CNNs. Interdependence between robustness and speed issues are also addressed. Another goal pursued is the unification of the theory of continuous-time and discrete-time systems. By means of a delta-operator approach, it is proven that the same network parameters can be used for both of these classes, even if their nonlinear output functions differ.;More complex CNN optimization problems that cannot be solved analytically necessitate resorting to numerical methods. Among these, stochastic optimization techniques such as genetic algorithms prove their usefulness, for example in image classificationproblems. Since the
暂无评论