Catálogo de publicaciones - libros

In the new H.264 video coding standard, motion estimation takes up a significant encoding time especially when using the straightforward full search algorithm (FS). We present an efficient scheme based on BP neural network algorithm which we believe can overcome to a significant degree this shortcoming. The mean squared error (MSE) between the current block and the same position in the reference frame is an often used matching criteria in block matching process. The scheme presented is very well suited to neural network training where the performance index is the mean squared error. The experimental results in Table 1 and Table 2 in the full paper compare our method with the full search algorithm. These comparisons show preliminarily but clearly that our method dose overcome to a significant degree the shortcoming of FS mentioned at the beginning of this abstract with neglectable coding efficiency loss.

Motion Estimation; mean squared error; BP neural network.

Image segmentation is of great importance in the field of image processing. A wide variety of approaches have been proposed for image segmentation. However, SAR image segmentation poses a difficult challenge owing to the high levels of speckle noise. In this paper, we proposed a SAR image segmentation method based on one-class support vector machines (SVM) to solve this problem. One-class SVM and two-class SVM for segmentation is discussed. One-class way is a kind of unsupervised learning, and one-class SVM based segmentation method reduces greatly human interactions, while yielding good segmentation results compared to two-class SVM based segmentation method. The segmentation results based on SVM are also compared to threshold method and adaptive threshold method. Experimental results demonstrate that the proposed method works well for image segmentation while reducing the speckle noise.

The adequacy of improved back propagation (IBP) neural network to model the inside air temperature and humidity of a production greenhouse as a function of outside parameters including temperature, relative humidity, wind speed, and solar radiation was addressed. To avoid standard BP algorithm’s shortcoming of trapping to a local optimum and to take advantage of the genetic algorithm (GA)’s globe optimal searching, a new kind of hybrid algorithm was formed based on the IBP neural network and GA. BP neural network was improved by adding the inertia impulse and self-adaptation learning rate to lessen convergence vibration and increase the learning speed. Then the initialized weights and thresholds of IBP neural network were optimized with GA. Through carrying out the experiments, the specimen data were collected on half-hourly basis in a greenhouse. After the network structure and parameters were determined reasonably, the network was trained. A comparison was made between measured and predicted values of temperature and relative humidity, and the results showed that the IBP neural network model combined with GA given a good prediction for inside temperature and humidity. By using the root mean square error (RMSE) algorithm, the RMSE between temperature predicted and measured was 0.8°C, and the relative humidity RMSE was 1.1%, which can satisfy with the demand of greenhouse climate environment control.

Aiming at the problem of parameter estimation in radar detection, a modified RBF neural network is proposed to estimate parameter accurately because of its good approximation ability to random nonlinear function and quick convergence speed. Two classical detection methods, which widely used in radar field, are listed in this paper, and their corresponding parameters are estimated with modified RBF neural network. Theoretical analysis and numerical results both show that the proposed method has good parameter estimation accuracy and quick convergence speed.

Optimal message scheduling is one of the key issues in the field of controller area network (CAN) bus system. There are numerous approaches related to this issue. Most of them are essentially based on priority-based strategies. In 1, we utilized Radial Basic Function (RBF) network 2 as a message scheduling controller to dynamically schedule messages. Furthermore, an online Backward-Through-Time (BTT) algorithm is presented for parameter optimization under fixed network structure. Intuitively, an inappropriate RBF network structure leads to performance degradation. In the worst case, the CAN system diverges. In this paper, we extend our previous works by including Minimal Resource Allocation (MRA) algorithm for structure determination. In this way, both problems of parameter optimization and structure determination can be resolved at the same time. Simulation results demonstrated that the proposed BTT with MRA methods outperform our previous results in terms of convergence time, stability, and the number of required hidden neurons (or radial basis functions).

Hardware/software partitioning of System-on-chip (SoC partitioning) has a significant effect on the cost and performance of the SoC. Given an embedded system specification and an available core library, the goal of low power SoC partitioning is to select appropriate intellectual-property (IP) cores or software components for the SoC, such that the power consumption of the SoC is minimized under price and timing constraints. SoC partitioning is first formulated to the constrained single-pair shortest-path problem in a directed, weighted graph, and then a novel discrete pulse coupled neural network (PCNN) approach is proposed to get the optimal solution. Autowaves in PCNN are designed specially to meet the constraints and find the optimal path in the constructed graph. Experimental results are given to demonstrate the feasibility and effectiveness of the proposed method.

In this paper, we address the problem of detection and correspondence of moving persons in a multi-camera set. A novel algorithm is proposed based on object occupancy random field (abbreviated by OORF), which has a robust performance even under severe occlusions and a fast implementation speed. The core of our algorithm is OORF and object window structure. The latter is essential to compute OORF and provides a scheme to detect and correspond objects simultaneously.

The paper describes the application of an artificial neural network in natural language text reasoning. The task of knowledge discovery in text from a database, represented with a database file consisting of sentences with similar meanings but different lexico-grammatical patterns, was solved with the application of neural networks which recognize the meaning of the text using designed training files. We propose a new method for natural language text reasoning that utilizes three-layer neural networks. The paper deals with recognition algorithms of text meaning from a selected source using an artificial neural network. In this paper we present that new method for natural language text reasoning and also describe our research and tests performed on the neural network.

This paper addresses the problem of designing a Markovian filter for a class of nonlinear stochastic Markovian jump systems. Firstly, neural networks are employed to approximate the nonlinearities in the different jump modes. Secondly, the overall system is represented by the mode-dependent linear difference inclusion (LDI). Then, a neural network-based Markovian filter is developed using the stochastic Lyapunov-Krasovskii stability theory under some linear matrix inequality (LMI) constraints. Finally, a numerical example is worked out to show the usefulness of the theoretical results.

This paper proposes a novel off-line signature verification method based on adaptive multi-resolution wavelet zero-crossing and one-class-one-network classification. First, the horizontal, vertical, 45 degree direction and the 135 degree direction projections of the binarizated signature images are calculated, respectively. The curvature data of the projections are decomposed into multi-resolution signals using wavelet transforms. Then the zero-crossings corresponding to the curvature data are extracted as features for verification. At last, one-class-one-network classifier is used to verify the signatures. The signature verification system was experimented on real data sets and the results show the system is very effective.