Catálogo de publicaciones - libros

The accuracy of 3D measurements of objects is highly affected by the errors originated from camera calibration. Therefore, camera calibration has been one of the most challenging research fields in the computer vision and photogrammetry recently. In this paper, an Artificial eural Network ased Camera Calibration ethod, , is proposed. The is especially useful for back-projection in the applications that do not require internal and external camera calibration parameters in addition to the expert knowledge. The offers solutions to various camera calibration problems such as calibrating cameras with automated active lenses that are often encountered in computer vision applications. The difference of the from the other artificial neural network based back-projection algorithms used in intelligent photogrammetry (photogrammetron) is its ability to support the multiple view geometry. In this paper, a comparison of the proposed method has been made with the undle lock djustment based back-projection algorithm, . The performance of accuracy and validity of the have been tested and verified over real images by extensive simulations.

A camera calibration method based on a nonlinear modeling function of an artificial neural network (ANN) is proposed in this paper. With the application of the nonlinear mapping feature of an ANN, the proposed method successfully finds the relationship between image coordinates without explicitly calculating all the camera parameters, including position, orientation, focal length, and lens distortion. Experiments on the estimation of 2-D coordinates of image world given 3-D space coordinates are performed. In comparison with Tsai’s two stage method, the proposed method reduced modeling errors by 11.45% on average.

A novel intelligent approach into 3D freeform surface reconstruction from planar sketches is proposed. A multilayer perceptron (MLP) neural network is employed to induce 3D freeform surfaces from planar freehand curves. Planar curves were used to represent the boundaries of a freeform surface patch. The curves were varied iteratively and sampled to produce training data to train and test the neural network. The obtained results demonstrate that the network successfully learned the inverse-projection map and correctly inferred the respective surfaces from fresh curves.

In volume data visualization, the classification is used to determine voxel visibility and is usually carried out by transfer functions that define a mapping between voxel value and color/opacity. The design of transfer functions is a key process in volume visualization applications. However, one transfer function that is suitable for a data set usually dose not suit others, so it is difficult and time-consuming for users to design new proper transfer function when the types of the studied data sets are changed. By introducing neural networks into the transfer function design, a general adaptive transfer function (GATF) is proposed in this paper. Experimental results showed that by using neural networks to guide the transfer function design, the robustness of volume rendering is promoted and the corresponding classification process is optimized.

In this paper, we present a novel real-time approach to synthesizing 3D character animations of required style by adjusting a few parameters or scratching mouse cursor. Our approach regards learning captured 3D human motions as parametric Gaussians by the self-organizing mixture network (SOMN). The learned model describes motions under the control of a vector variable called the , and acts as a probabilistic mapping from the low-dimensional style values to high-dimensional 3D poses. We have designed a pose synthesis algorithm and developed a user friendly graphical interface to allow the users, especially animators, to easily generate poses by giving style values. We have also designed a method, which accepts a sparse sequence of key style values and interpolates it and generates a dense sequence of style values for synthesizing a segment of animation. This method is able to produce animations that are more realistic and natural-looking than those synthesized by the traditional key-framing technique.

This paper focuses on enhancing the effectiveness of filter feature selection models from two aspects. One is to modify feature searching engines based on optimization theory, and the other is to improve the regularization capability using point injection techniques. The second topic is undoubtedly important in the situations where overfitting is likely to be met, for example, the ones with only small sample sets available. Synthetic and real data are used to demonstrate the contribution of our proposed strategies.

In the field of neural networks, feature selection has been studied for the last ten years and classical as well as original methods have been employed. This paper reviews the efficiency of four approaches to do a driven forward features selection on neural networks . We assess the efficiency of these methods compare to the simple Pearson criterion in case of a regression problem.

The unlabeled document or text collections are becoming larger and larger which is common and obvious; mining such data sets are a challenging task. Using the simple word-document frequency matrix as feature space the mining process is becoming more complex. The text documents are often represented as high dimensional about few thousand sparse vectors with sparsity about 95 to 99% which significantly affects the efficiency and the results of the mining process. In this paper, we propose the two-stage Non-negative Matrix Factorization (NMF): in the first stage we tried to extract the uncorrelated basis probabilistic document feature vectors by significantly reducing the dimension of the feature vectors of the word-document frequency from few thousand to few hundred, and in the second stage for clustering or classification. In our propose approach it has been observed that the clustering or classification performance with more than 98.5% accuracy. The dimension reduction and classification performance has observed for the Classic3 dataset.

Pulse coupled neural network (PCNN) is different from traditional artificial neural networks, models of which have biological background and are based on the experimental observations of synchronous pulse bursts in the cat visual cortex. However, it is very difficult to determine the exact relationship between the parameters of PCNN model. Focusing on the famous difficult problem of PCNN, how to determine the optimum parameters automatically, this paper proposes the definition of water valley area, establishes a modified PCNN, and puts forward an adaptive PCNN parameters determination algorithm based on water valley area. Extensive experimental results on image processing demonstrate its validity and robustness.

Multirate systems are abundant in process industry, many soft-sensor design problems are related to modeling, parameter identification, or state estimation involving multirate systems. In this paper, a polynomial transformation technique has been used to derive a dual-rate model with a finite number of parameters; based on this model, the dual-rate forgetting gradient algorithm has been used to estimate the model parameters and intersample outputs based on the dual-rate input-output data directly. Furthermore, convergence properties of the algorithms in the stochastic framework are studied and show that 1) the parameter estimation error consistently converges to zero under the persistent excitation condition; 2) the intersample output estimation error is uniformly bounded. Finally, a simulation example show excellent effectiveness in parameter and output estimation.