This paper proposed an algorithm that extracts the optimum characteristics parameters to identify the signatures from the spectrum using 2-D FFT. The signature image input through a scanner is normalized into 250*128 pixels in the prepocessor. Normalized image is divided into block segments and each segment is transformed into space spectrum by 2-d FFT. There are several methods extracting the signature characteristic parameters from that spectrum. The result of experimentations which use the characteristic parameters extracted between $0^{\circ}and\;90^{\circ}$ in (0, 0), (63, 0) corners from 64$\times$64 block spectrum shows that the signature identification rate using that method gives 92.5% of successful achievement for 100 signatures, higher than the others.

This paper describes a normalization of character which is contained in the character recognition process. Line and dot density is computed on input character image and then image mapping is executed into destination. Also recognition is processed using overlap-partitioning of character image and extraction of 4 directional feature primitives. The validity of proposed nonlinear normalization algorithm could be verified by increment of recognition rate.

As a method of separating the signal arriving from a desired direction in the presence of noise and interfering signals, a linearly constrained adaptive beamformer based on the escalator (Gram-Schmidt) structure is presented. The weights are chosen to minimize local output power subject to the response constraint. The performances of the proposed escalator adaptive beamformer are compared with those of cascade adaptive beamformer via computer simulations.

This paper presents a decomposition method to evaluate the performance measures of the serial production lines with unreliable machines and finite buffers. This method is to decompose the transfer line into a set of two machine lines for analysis. Bases on this approximation method, we consider nonhomogeneous lines. In such lines, the machines may take the different lengths of time performing operations on parts. A new transformation is employed in order to replace the line by a homogeneous line. The approximate transformation enables one to determine parameters of performance such as production rate and average buffer levels, and it also shows better application for a large class of systems.

This paper presents a hardware for image alignment based on proposed new algorithm which can align a small misaligned document image simply by one transformation with a parallel shifting of pixels. This hardware is simulated with VHDL and estimated to be about 65 ms to align an image made up of 380 by 480 pixels. Also, we will demonstrate the effectiveness of the proposed image alignment algorithm in OCR system by comparing its characteristics with those of the existing image rotation algorithms.

The tasks which detect and track moving objects, by analyzing dynamic images taken at a constant time interval, are essential in various applications. This paper suggests how to utilize domain-specific knowledge and motional knowledge for detecting and tracking moving objects. That is, The trajectory information of a moving object is to be used for generating hypotheses on expected motion and expected position of moving objects, and the domain-specific knowledge is to be used for verifying the generated hypotheses.

This paper proposes a method for improving the performances of the neural network for optimization by an optimal estimation of initial states. The optimal initial state that leads to the global minimum is estimated by using the stochastic approximation. And then the update rule of Hopfield model, which is the high speed deterministic algorithm using the steepest descent rule, is applied to speed up the optimization. The proposed method has been applied to the tavelling salesman problems and an optimal task partition problems to evaluate the performances. The simulation results show that the convergence speed of the proposed method is higher than conventinal Hopfield model. Abe's method and Boltzmann machine with random initial neuron output setting, and the convergence rate to the global minimum is guaranteed with probability of 1. The proposed method gives better result as the problem size increases where it is more difficult for the randomized initial setting to give a good convergence.

This paper presents an efficient SOFM(self-organizing feature map) algorithm for the solution of the large scale TSPs(traveling salesman problems). Because no additional winner neuron for each city is created in the next competition, the proposed algorithm requires just only the N output neurons and 2N connections, which are fixed during the whole process, for N-city TSP, and it does not requires any extra algorithm of creation of deletion of the neurons. And due to direct exploitation of the output potential in adaptively controlling the neighborhood, the proposed algorithm can obtain higher convergence rate to the suboptimal solutions. Simulation results show about 30% faster convergence and better solution than the conventional algorithm for solving the 30-city TSP and even for the large scale of 1000-city TSPs.

This paper presents a neural network IC for Korean vowels recognition. The neural network is composed with three levels and which is learned by Back Propagation algorithm. In the neural network IC, the neuron bodys and synapses are implemented with CMOS inverters and ion-implanted polysilicon resistors.