Iulian-Constantin Vizitiu

Optimal FPGA implementation of GAMLP systems

An interesting approach to assure the real-time property of neural automatic target recognition systems is to use an efficient hardware implementation of the neural networks used inside of their classification chains. Consequently, a proper genetic procedure used to optimize both connectivity and distribution of the neural weights of MLP neural networks (GAMLP system) is presented. Finally, having as starting point the previous broached aspects, an optimal FPGA hardware implementation of MLP neural networks is also described.

Sidelobe reduction in pulse-compression radar using the stationary phase technique: An extended comparative study

According to pulse-compression radar theory, the sidelobe reduction using nonlinear frequency modulation (NLFM) signal processing represents a major and present research direction. Although in literature a lot of techniques to design efficient NLFM waveforms are indicated, one of the most important methods as application area is focused on stationary phase technique (SPT) use. Consequently, the main objective of this paper is to make an extended comparative study as sidelobe suppression among some very promising NLFM laws achieved by applying of a proper SPT synthesis algorithm. In addition, some aspects related to the suitable choice of the parameters involved in the synthesis process of NLFM signals are also indicated.

High-quality HRR ATR system using an improved neural recognition chain

One of the most recent technique to design an efficient ATR system is to use high-resolution radar (HRR) imagery as input information flow. To increase the quality of such system, an interesting approach is to use powerful artificial neural networks inside of its recognition chain. Consequently, an improved neural recognition function based on modified feature extraction and selection methods and respectively, on genetic optimized RBF network architecture is described. Finally, to confirm the broached theoretical aspects, a real HRR image database was also used.

High-performance pattern recognition system using an improved neural classification chain and decision fusion on multispectral information

The performance level of pattern recognition (PR) systems can be improved using in a proper way new powerful artificial neural networks inside of its classification chains and respectively, decision fusion techniques on available multispectral information. Consequently, to increase the pattern recognition performances, an improved neural classification chain and respectively, an evolutive version of Sugenos fuzzy integral based on high-resolution radar (HRR), video and thermal imagery use are described. To confirm the broached theoretical aspects, a real image database was used.

An optimal full-genetic technique used to train RBF neural networks

It is well-known that, the pattern recognition performances assigned to RBF neural networks depends a lot by their specific training algorithms, and by the methods used for RBF center selection (e.g., a clustering technique), particularly. Having as starting point the membership of genetic algorithms to the powerful class of global optimization methods, an optimal full-genetic training procedure of RBF neural networks based on hybrid genetic clustering algorithm used for center mapping, and on genetic approach to fit the output neural weights is proposed. Finally, using a real pattern recognition task, a comparative study (as performance level) with others standard RBF training methods and SART neural network is also described.

GANN system to optimize both topology and neural weights of a feedforward neural network

An interesting approach to improve the quality of the artificial neural network architectures included into a large spectrum of applications, is to use the GANN (Genetic Algorithm Neural Network) system concept. Consequently, a specific genetic technique which simultaneously optimizes both topology and neural weights of a feedforward neural network is described. Finally, to confirm the broached theoretical aspects, a real training database was also used.

More efficient ATR system using the decision fusion between HRR and video imaginary

One of the most important methods to improve automatic target recognition (ATR) function is to use fusion techniques. In this paper we propose an application using an improved version of Sugenopsilas fuzzy integral to increase the target recognition performance based on high-resolution radar (HRR) and video imaginary. In order to confirm the broached theoretical aspects, a real database was used.

A new invariant set for video pattern recognition in ATR systems

The choice of an appropriate feature extraction method is essential for the success of the classification or recognition process. T his paper proposes a design method for a new pattern descriptor set based on the Flusser moment class, which is invariant to elementary geometric transformations and has an increased robustness to the action of some perturbations. Experimental results based on the use of a real video image database confirm the basic properties of this new descriptor set.