Minoru Fukumi

Rotation-invariant neural pattern recognition system with application to coin recognition

In pattern recognition, it is often necessary to deal with problems to classify a transformed pattern. A neural pattern recognition system which is insensitive to rotation of input pattern by various degrees is proposed. The system consists of a fixed invariance network with many slabs and a trainable multilayered network. The system was used in a rotation-invariant coin recognition problem to distinguish between a 500 yen coin and a 500 won coin. The results show that the approach works well for variable rotation pattern recognition.

Rotation-invariant neural pattern recognition system estimating a rotation angle

A rotation-invariant neural pattern recognition system, which can recognize a rotated pattern and estimate its rotation angle, is considered. It is well-known that humans sometimes recognize a rotated form by means of mental rotation. The occurrence of mental rotation can be explained in terms of the theory of information types. Therefore, we first examine the applicability of the theory to a rotation-invariant neural pattern recognition system. Next, we present a rotation-invariant neural network which can estimate a rotation angle. The neural network consists of a preprocessing network to detect the edge features of input patterns and a trainable multilayered network. Furthermore, a rotation-invariant neural pattern recognition system which includes the rotation-invariant neural network is proposed. This system is constructed on the basis of the above-mentioned theory. Finally, it is shown that, by means of computer simulations of a binary pattern and a coin recognition problem, the system is able to recognize rotated patterns and estimate their rotation angle.

Recognition of EMG signal patterns by neural networks

The paper tries to recognize EMG signals by using neural networks. The electrodes under the dry state are attached to wrists and then EMG is measured. These EMG signals are classified into seven categories, such as neutral, up and down, right and left, wrist to inside, wrist to outside by using a neural network. The neural network learns FFT spectra to classify them. Moreover, we perform the principal component analysis using the simple principal component analysis before we perform recognition experiments. It is shown that our approach is effective to classify the EMG signals by means of computer simulations.

Face recognition using genetic algorithm based template matching

We present a face recognition method using template matching. Template matching is performed with the help of a genetic algorithm to automatically test several positions around the target and also adjust the size of the template as the matching process progresses. We use two kinds of templates for each face. One is based on edge detection and the other depends on the YIQ color information from the face. Our template is a T-shaped region symmetrical between the eyes, covering both eyes, the nose and mouth. Our features of interest to achieve face recognition are therefore the eyes, nose and the mouth. We ignore the shape of the face so as to have a small template for faster matching and also because the effect of the shape does not result in a significant increase in the overall final accuracy. We conducted a simulation experiment to verify our idea and also did a comparative experiment using a distance measure face recognition method.

Face detection based on skin color information in visual scenes by neural networks

A method to examine whether or not human faces are included in the images and to detect their position by using the technique of skin color region extraction is presented. In this technique, the skin color which is a main feature of faces is detected, a binary image composed of skin color parts and background one is constructed from an original image using a neural network which learns color information, and then the skin color parts of some sizes are regarded as face candidates. Thus search regions are limited within the skin color parts. Therefore, an improvement in the detection speed is achieved. These face candidates are examined using a neural network which learns the features of faces, and estimates whether or not the original image includes the faces. From results of computer simulations, a search rate of 83.3 % accuracy was achieved from 15 sheets, each having from 1 to 3 faces. The sizes and positions of faces were chosen as randomly as possible. There was no search of other objects other than faces.

Apparent age estimation system based on age perception

The age is one of important information in our living. If the age estimation that uses face image by computer becomes possible, it is thought that the age estimation assumes an important role in various scenes. In this paper, we propose an age estimation every age by using the supervised SOM. Furthermore, the important features for the age estimation are selected by the GA.

3-D eddy current calculations using the magnetic vector potential

Two methods for using the magnetic vector potential for 3-D eddy current calculation are treated. One method uses the magnetic vector potential that is continuous over the entire region and generally accompanies the electric scalar potential. It has the advantage that no cutting is necessary for the multiply-connected-region problem. The other method uses the magnetic vector potential that is discontinuous across the interface surface between different media. This magnetic vector potential can be arranged so that the electric scalar potential does not appear in the equations when the conductivity is constant. It has the disadvantage that cutting is necessary for the multiply-connected-region problem. New boundary value problem formulations are given for both methods, precisely defining the interface and boundary conditions. >

License plate detection system by using threshold function and improved template matching method

License plate recognition is very important in an automobile society. However, it is very difficult to do it, because a background and a car surface color can be similar to that of the license plate. Furthermore, detection of cars moving at a very high-speed is difficult to be done. We propose a new method to extract a car license plate automatically by using a genetic algorithm (GA). By using GA, the most likely plate colors are decided under various light conditions. First, the average brightness Y values of images are calculated. Next, relationship between the Y value and the most likely plate color thresholds (upper and lower bounds) are obtained by GA to estimate threshold equations by using the RLS algorithm. Finally, in order to show the effectiveness of the proposed method, we show simulation examples by using real images.

Designing a neural network for coin recognition by a genetic algorithm

This paper presents a method to design a neural network for coin recognition by a genetic algorithm (GA). The GA specifies an architecture of neural network, but does not train the network. The back-propagation (BP) method trains the network. After training it by the BP, the GA varies the architecture of the network to fit the environment, which is to achieve a 100% recognition accuracy and to make the network small in size. The network reduced by the GA is further decreased by using the BP with forgetting of weight. The object of this paper is to design a smaller neural network for hardware implementation of coin recognition system. Results by computer simulation show the effectiveness of the method to variably rotated coin recognition problem.

A new back-propagation algorithm with coupled neuron

Summary form only given, as follows. A novel algorithm is developed for training multilayer fully connected feedforward networks of coupled neurons with both signoid and signum functions. Such networks can be trained by the familiar backpropagation algorithm since the coupled neuron (CONE) proposed uses the differentiable sigmoid function for its trainability. The algorithm is called CNR, or coupled neuron rule. The backpropagation (BP) and MRII algorithms which have both advantages and disadvantages have been developed earlier. The CONE takes advantages of the key ideas of both methods. By applying CNR to a simple network, it is shown that the convergence of the output error is much faster than that of the BP method when the variable learning rate is used. Finally, simulation results illustrate the effective learning algorithm.<<ETX>>