This paper proposes a writer identification and verification method using autoassociative neural networks. An autoassociative neural network is separately prepared for each category and doesn't depend on any other category in its learning processing, so it is casily applicable to the increase of categories to be distinguished. Weighted direction index histogrzam feature of 256 dimensions extracted from the handwritten character patterun is used as an input value to this network. In the network, it is ideal when the output value will be equal to the input value. Therefore, writer identification is executed by compairing the errors between input and output values of every network, and writer verification is performed by compairing the error in the network with previously assigned threshold value. From the experiment applied to 20 persons, the correct identification rate of 92.48% was obtained and the correct acceptance rate was 92.73% in the writer verification. Furthermore, it is clarified that the high precision writer recognition will be feasible by using the combined character patterns of the same category or the different category in the calculation of errors.
This paper proposes a character segmentation and spotting method of historical documents. In the segmentation method, the result of character recognition process is utilized to cope with the cursive scripts and the mutual encroachment of characters which are peculiar to the historical documents. In the spotting method, the previously designated characters are only extracted from the characters string. As an early segmentation, the characters string pattern is divided into the same connected component by using the labelling processing. The area composed of the same component is surrounded with a rectangle and each character pattern is segmented each other by using the shape of rectangle such as height and width. Next, the individual character recognition is applied to the segmented pattern. From the recognition result, the rectangle failed in the segmentation is picked up and the resegmentation is applied to the string contains this rectangle. Therefore, it is expected that the string is divided at the best position. On the other hand the neural network which corresponds to the previously designated character is prepared. The error between input and output of the network applied to the segmented pattern is calculated and the pattern which satisfies the condition is extracted as a spotting result. From the extraction experiment applied to 615 characters strings, the correct spotting rate of 94.22% was obtained to 5 designated characters by using the resegmentation process, but the rate was 87.58% without the resegmentation process.
Although the Hopfield neural network is suitable for hardware implementation, local minimum convergence is one of the most serious problems which must be solved by using simple mechanism. Hardware implemention of the model contributes to a fast solver of the combinatorial optimization problems. The maximum neural network has simple structure and can easily find a feasible solution. In this paper, the maximum neural network is modified by employing reinforced self-feedbacks and logical synaptic connections in order to implement by using simple digital circuits. The proposed model is applied to the N-queens problem and the multi-layer channel routing problem in VLSI automation design. From numerical experiments it is confirmed that the maximum neural network with reinforced self-feedbacks and synaptic connections is more effective than the conventional models. A detail architecture of the model for FPGA implementation is shown, and flexibility for various problems is clarified. Lastly the proposed model is implemented on Xilinx FPGA XCV200 and scalability is discussed.
Discharge characteristics have been investigated for a needle-plane electrode configuration containing insulating barriers with a narrow gap which has been placed between the needle and the plane. Characteristics of creeping discharge developed in the narrow gap of the barrier filled with SF6 have been focusd. In the case of a configuration with a backside electrode below the needle, corona generated from the needle has easily extended to the gap. On theother hand, for the case of configuration without backside electrode, the corona has hardly extended to the gap, and then by increasing applied voltage the corona has greatly developed in the gap. This difference on the corona extension should affect to the flashover characteristics in the present system.
This paper presents a method for fast off-line logical location of faults in analog electronic circuits at the sub-network level and verifies its practical diagnosability. The proposed approach breaks through the previous limitation that all torn terminals (incident nodes) must be accessible and that the mutual-testing method must be utilized to locate the faulty sub-networks. As far as the diagnosability is concerned, its application is more extensive than the unified decomposition approach. Therefore it better satisfies the engineering needs.
Recently, a new type of bifurcation phenomena called Border-Collision (abbr. BC) bifurcation has been discovered for one or two-dimensional one-parameter families of discrete-time systems. A skewed tent map is the simplest model exhibiting the BC bifurcation and the existence of break point (undifferentiable point) in the map is an essential for BC bifurcation. In some continuous systems, BC bifurcation is also confirmed. In particular, many interesting bifurcation phenomena including BC bifurcation are observed in the power electronic systems. Since power electronic circuits with current or voltage feedback have wide industrial applications, indeed, theoretical and experimental analysis in such systems are very important in practical point of view. In this paper, we propose a system interrupted by own state and a periodic interval, and show that the system has much variety of bifurcation phenomena, including BC bifurcation. To analyze properties of the dynamics, we derive a one-dimensional map explicitly. We show some theorem and the regions of periodic solution within two parameter space. Some of theoretical results are verified by laboratory experiments.
An endoscope image has a characteristic that the image is presented by reddish colors, since much capillary vessels are located in internal walls of human organs. In the method, the image information is compressed by a characteristic of reddish endoscope image. Ordinary techniques like as DCT, DPCM are not used in the method. Compressed information is constituted by a header information section and a image construction information section. In the header information section, color numbers of each pixel are indicated by difference between its neighbor. In the image construction information section, the color number appearance rate is indicated by index symbols. The index symbols are defined as small number like 0-0, 0-1, 0-2, … in a matrix. In the experimental result, the compression rate of proposed method is not better than the case of loss-less JPEG process in large color region like a full-scaled typical endoscope images. However, the rate of the proposed method rises in small color region. Under approx. 20000 colors, the rate is better than the loss-less JPEG process. In a case of ROI#1 which is constructed by approx. 1700 colors, difference between compression rate of proposed method and the rate of lossy JPEG is less than 4%. It is considered that the proposed method is effective to compress the small color number endoscope image like as ROI image.
The mutual coupling between antenna elements usually contributes a significant effect to the performance of adaptive array, especially when the interelement spacing is small. In this paper, we discuss the edge effects on a pattern synthesis based on an adaptive array theory. To diminish the edge effect of adaptive array, we simulate the reduction of mutual coupling by adding shorted parasitic elements. We can compensate the edge effects by adding shorted parasitic elements. The algorithm can be used on a large class of design problems.
We proposed the recognition system of road signpost by image processing. At first, in the preprocessing, the red color is extracted from the input image using color chart. Secondly, the ring detection network specifies the scale and position of the circular ring by scanning the input image. The figure pattern is cut out and reconstructed using interpolation method, assumed to consist of 8 length and 5 dots. A digital expression of the template and the image is changed -1, 0, +1. set by the experiment, and we experiment on the performance of ring detection network based on it. The recognition experiment on 100 pieces was done for three kinds of speed limit signs 30 km/h, 40 km/h, and 50 km/h. The detection rate was about 100% in all cases. The recognition rate was 98.3%. This network is possible to detect the other types of shape, using other template. The application can be expected of many fields of the image processing. Moreover, these processes are carried out using networks by many connecting weights. These results suggests in apprication of parallel hierarchical structure in hardware system, and is able to realize real time processing.
A modeling and control method for a class of nonlinear systems whose dynamics depend on a process variable is presented. Two ExpARMAX (exponential ARMAX) models, a type of global NARMAX (nonlinear ARMAX) model are used to develop a multi-step predictive control algorithm which does not use on-line parameter estimation for the nonlinear system. One ExpARMAX model is used as the internal model of the predictive controller, and the other is used to predict future values of the measurable disturbances used in the predictive controller. The global ExpARMAX models only require off-line identification, and their local linearization forms are similar to linear ARMAX models. Case studies taken from a selective catalytic reduction (SCR) process for the reduction of Nitrogen Oxide (NOx) emissions of thermal power plants verify the effectiveness of the method.
Self-tuning control schemes (STC) are useful for systems with unknown or slowly time-varying parameters. Some single-input/single-output PID control schemes based on STCs have been proposed for such systems. However, there are a lot of multivariable systems in real process industries. And these systems often have relatively large time-delays. In this paper, a design scheme of self-tuning PID control system is proposed for multivariable systems with unknown parameters and time-delays. The controlled object is equipped with an internal model in order to compensate the time-delay and also unstable zeros. Subsequently, a multivariable PID controller is designed for the augmented or compensated system. The PID parameters are calculated recursively based on the relationship between the mimimum variance control law and the PID control law. A simulation example is presented to demonstrate the effectiveness of the proposed scheme.
Dynamic uniform temperature on thermal conduction surface is widely required in many applications. We proposed a control method using the gradient of temperature between the arbitrary points. This paper presents the experimental which verifies the effectiveness of the proposed control method for the dynamic uniform temperature on thermal conduction surface.
This paper describes the practical application of robot vision techniques to electricity distribution works. To cope with difficult outdoor lighting conditions together with various background scene caused by a wide variation of the working environment structures, we have developed two methods of detection of an object using both 2-D gray-scale image and 3-D range image. The first method finds an object of interest in a 2-D image taken from a rather far position. Although a 3-D range image taken from the same position has too low resolution to extract surface geometry of the object, it is effectively used to eliminate the background scene from the 2-D image based on the distance threshold. The second method generates online template image of an object to be manipulated. This enables reliable real time visual tracking by stereo vision in the presence of variation of shape and appearance of the object. We show how these methods are implemented and effectively used in the distribution work.
Neural networks with input gates are proposed for behavior learning of agents. The networks are equipped with gates on their input channels that pass input signals when they are neccesary. A gate opens and closes depending on the current values of input signals. The dependence is automatically determined based on training data. They are applied to behavior learning of agents in the reinforcement learning framework. The gate openings provide the generalized information about the significance of each input signal, which reduces the size of region to be explored and can be exploited to speed up the subsequent learning in other environments.
In this article, we investigated possibility to establish a recycling-based industrial system in Japan. For this purpose, a novel mathematical model was first developed for the recycling-based system by expanding input-output analysis. Utilizing this model, we clarified several significant implications on the possibility of the recycling-based system and the prices of recycled goods. Then various recycling technologies were evaluated to clarify how these technologies could contribute to establish the system as well as to mitigate global warming. Next we developed an optimization model of the entire industrial systems in Japan, including conventional and emerging technologies on waste disposal and recycling. Various institutions on the recycling system were investigated utilizing this model. These include prices of recycled goods and tradeoff relationships between promotion of recycling and reduction of CO2 emissions. In particular, evaluated results showed that shadow price of recycled goods did not necessarily become positive. We also proposed a novel institution on the recycling system as “Waste dumping permit” in Japan. In conclusion, we clarified a desirable strategy so as to establish the recycling-based system, as well as to mitigate global warming.
Maximum cut is an important of a class of combinatorial optimization problems. It has many important applications including the design of VLSI circuits and design of communication networks. The goal of this NP-complete problem is to partition the node set of an undirected graph into two parts in order to maximize the cardinality of the set of edges cut by the partition. In this paper, we propose a parallel algorithm using gradient ascent learning algorithm of the Hopfield neural networks for efficiently solving such optimization problems. The proposed learning algorithm is tested on a 2-variable quadratic polynomial and applied to the MAX CUT problem. Extensive simulations are performed and its effectiveness is confirmed.