Naoki Oshiro

Binocular tracking using log polar mapping

This paper describes a new binocular tracking method using log polar mapping (LPM) which approximately represents the mapping of the retina into the visual cortex in primate vision. Using LPM makes it possible not only to obtain both a high central resolution and a wide field of view, but also to significantly reduce the processing of image data. In this paper, LPM is performed in software by the lookup table method. Our tracking method utilizes a zero disparity filter (ZDF) for extracting the target object and virtual horopter method for estimating binocular disparities, respectively. The performance of both target extraction and disparity estimation is improved in comparison with the conventional methods, by using LPM. Some experimental results are also shown to demonstrate the effectiveness of the proposed method.

Searching performance of a real-coded genetic algorithm using biased probability distribution functions and mutation

One excellent crossover method for the real-coded genetic algorithm (RGA) is the unimodal normal distribution crossover method (UNDX). The UNDX is superior to the blend crossover method (BLX). The UNDX uses Gaussian distribution functions based on the main and sub searching lines. In this article, we present a method of improving the searching performance of the RGA. We propose the use of biased probability distribution functions (BPDFs) based on the main and sub searching lines in the crossover process. The crossover with BPDFs frequently produces offspring that are close to the best individuals in the current generation, and it is highly likely that these offspring will offer the best solution to the problem. Furthermore, we propose a mutation that has a constant and extended range that is wider than that of the UNDX. Simulations show the efficiency of the proposed method.

Foveated Vision for Scene Exploration

In this paper, foveated vision for scene exploration is implemented. The peripheral and central vision are the basic capabilities of foveated vision. The informations obtained from the peripheral vision are used to determine the next gaze point. Due to the low resolution of the periphery, however, the determination is not always appropriate. To solve this problem, we propose to evaluate the target object by the central vision after gazing. We implement foveated vision based on the Log Polar Mapping (LPM) and construct an evaluation scheme of the target object in the central vision using LPM rotational-invariance. The peripheral vision is realized by Zero Disparity Filter for LPM stereo images. Some experimental results are also shown to demonstrate the effectiveness of the proposed method.

Separating visual information into position and direction by SOM

A model is proposed to self-organize a map for the visual recognition of position and direction by a robot moving autonomously in a room. The robot is assumed to have visual sensors. The model is based on Kohonen’s self-organizing map (SOM), which was proposed as a model of self-organization of the cortex. An ordinary SOM consists of a two-dimensional array of neuron-like feature detector units. In our model, however, units are arranged in a three-dimensional array, and a periodic boundary condition is assumed in one dimension. Also, some new learning rules are added. Our model is shown by a computer simulation to form a map which can extract from the visual input two factors of information separately, i.e., the position and direction of the robot. This is an example of so-called two-factor problems. In our algorithm, the difference in the topology of the information is used to separate two factors of information.

Separating visual information into position and direction by two inhibitory connected SOMs

In this article, we propose a model to self-organize a map for robot navigation using its own visual information. The robot is assumed to have visual sensors around it. The recognition model is based on Kohonen’s self-organizing map (SOM), which was proposed as a model of the self-organization of a cortex. An ordinary SOM consists of a two-dimensional array of neuron-like feature detector units. We want to extract the direction and position information separately from the visual input, which is a function of the two information factors. Our model consists of two layers. The first layer is for directional information, and consists of units arranged in a circular array, and the second layer is for position information, and consists of a two-dimensional array. The units in the second layer accept inputs from all the units in the first layer through plastic inhibitory synapses. It will be shown by computer simulation that the units in the first layer develop direction sensitivity and lose position sensitivity through training, while in the second layer, the units develop position sensitivity and lose direction sensitivity.

Backward movement control with two-trailer truck system using genetic programming

In this paper, we propose a control system using genetic programming (GP) for backward movement control of a two-trailer truck, known as a nonholonomic system. We have already achieved the control of a single trailer using GP. In this study, we aim to design a control system for complex problem of two trailers. In order to verify the effectiveness of the proposed method, it is compared to neuro controller (NC) system evolved by genetic algorithm (GA).

A self-organizing VQ model of head-direction cells and grid cells

In this article, we propose a three-layered self-organizing model to extract the head direction and position of a moving object separately. The model consists of three layers, each of which is a self-organizing vector quantization (VQ) model. The second layer receives inhibitory input from the first, and the third layer receives inhibitory input from the first and second. The first layer is to detect the head direction, and the second and third are to detect the position. The information representation in the second and third layers is shown to be a multi-ary expression, and the units in the third layer develop a receptive field with a grid structure, as was observed in the entorhinal cortex of a rat.

A self-organizing model of place cells with grid-structured receptive fields

In this article, we propose a new information separation model consisting of two vector quantization (VQ) layers, the superior layer and the inferior. The inferior VQ layer receives inhibitory input from the superior via anti-Hebbian synapses, which forces the winner on the inferior layer to be distributed independently of that on the superior. Supplied with input vectors carrying 2D positional information, the inferior layer can self-organize place-cell-like units with a grid-structured receptive field which shows a remarkable resemblance to that of “grid cells” found recently in the entorhinal cortex of a rat.

Investigation of multi-layer neural network performance evolved by genetic algorithms

This paper presents a study on the investigation of multi-layer neural networks (MLNNs) performance evolved with genetic algorithm (GA) for multi-logic training patterns applied to various network functions. Specifically, we have concentrated on the Sigmoid, Step and ReLU functions to evaluate and simulate their performances in the network. We have revealed that GA training gives good training results in evolutionary computation by changing of Sigmoid, ReLU and Step as the activity functions in MLNN performance. Sigmoid function has proved to train all patterns for all outputs without any challenge as compared to ReLU function and Step in this study. We are still trying to see how a ReLU function could be trained with GA for MLNNs performance for the two input and four output training patterns termed as the multi-logic pattern training about multiple training parameters.

Fruit maturity detection using neural network and an odor sensor: Toward a quick detection

Maturity detection is very important for fruit farmhouses. In a previous study, we revealed a type of odor sensor that responds to the strength of the fruits smell as well as to the fruits maturities. The smell data consists of a dead time and a step response of a first-order lag element. We focus on the step response of first-order lag element, which is a form that rises exponentially to a constant value. This paper presents a quick detection method of fruit maturity in a few seconds of the rising signal of the odor sensor. Using neural network, the method performs without waiting for the sensor to fully reach up to a constant value. First, a neural network is trained for sample data with two kinds of maturities: fully ripe and immature. By testing the neural network with untrained data, we confirmed that the network is able to detect the fully-ripened, middle-ripened, and unripe fruits.