矢田 紀子

人間の図形認識メカニズムは，単純な図形の組み合わせで複雑な図形を認識する機構であると考えられている．これが "図形アルファベット仮説" である．我々はこの考え方を "比較的単純な図形群の中から選択された図形の組み合わせによって複雑な図形を認識する機構" ととらえ，コンピュータによるパターン分類への応用を提案する．提案手法では，単純な図形と仮定した N×N 画素のドットパターン （Alphabet Dot Pattern；ADP） と認識対象図形のハミング距離を特徴量として分類する．本論文では，まず提案手法のシステムについて述べる．また，実験の一例として提案手法をマルチフォント図形の認識に適用した結果について示す．この結果，提案手法では比較手法よりも高い分類正答率が得られた．It is thought that human being recognizes a complicated figure by combining simple figure.This is "figure alphabet hypothesis" and these simple figures are called a figure alphabet.We considered "the mechanism in which a complicated figure is recognized with the combination of the figure chosen from comparatively simple figure groups", and applies it to a pattern classification. The proposed method assumes the figure alphabet to be the dot pattern (Alphabet Dot Pattern；ADP) of an N×N pixel. Because there are many kinds of ADP, ADP group is optimized by the genetic algorithm (GA).And, the hamming distance of an input figure and an ADP group is calculated, and classifies a figure. In this research, the proposed method was applied to recognition of a multifont figure as an example of an experiment, and the validity of the proposed method was verified.Consequently, the result of the proposed method was able to obtain the correct answer rate higher than the result of a comparative experiment. This paper describes the way of thinking and experiment result of the proposed method.

近年，画像を身近に扱う機会が増えたことで，計算機による画像分類のニーズがより一層高まっている．Genetic Image Network for Image Classification (GIN-IC) は進化計算によって自動的に画像変換部を含む画像分類アルゴリズムを構築する手法であり，その有効性が示されている．本報告の目的は，AdaBoost のアルゴリズムに基づいて GIN-IC の性能を向上させることである．具体的には GIN-IC を弱識別器として利用し，相互補完させることで性能の向上を図る．実験では，提案手法を 3 種類の画像分類問題に適用した結果を示す．これらの実験から，提案手法の学習および検証用画像に対する分類精度が，従来の GIN-IC と比較して改善されることが確認できた．Automatic construction method for image classification algorithms have been required. Genetic Image Network for Image Classification (GIN-IC) is the automatic construction method for image classification algorithms which include image transformation component using evolutionary computation, and its effectiveness has already been proven. In our study, we try to improve the performance of GIN-IC with AdaBoost algorithm using GIN-IC as weak classifiers to complement with each other. We apply our proposed method to three types of image classification problems, and show the results in this paper. In our method, discrimination rates for training images and test images improved in the experiments compared with the previous method GIN-IC.

人間の図形認識メカニズムは,単純な図形の組み合わせで複雑な図形を認識する機構であると考えられている.これが "図形アルファベット仮説" である.我々はこの考え方を "比較的単純な図形群の中から選択された図形の組み合わせによって複雑な図形を認識する機構" ととらえ,コンピュータによるパターン分類への応用を提案する.提案手法では,単純な図形と仮定した N×N 画素のドットパターン (Alphabet Dot Pattern;ADP) と認識対象図形のハミング距離を特徴量として分類する.本論文では,まず提案手法のシステムについて述べる.また,実験の一例として提案手法をマルチフォント図形の認識に適用した結果について示す.この結果,提案手法では比較手法よりも高い分類正答率が得られた.

近年,画像を身近に扱う機会が増えたことで,計算機による画像分類のニーズがより一層高まっている.Genetic Image Network for Image Classification (GIN-IC) は進化計算によって自動的に画像変換部を含む画像分類アルゴリズムを構築する手法であり,その有効性が示されている.本報告の目的は,AdaBoost のアルゴリズムに基づいて GIN-IC の性能を向上させることである.具体的には GIN-IC を弱識別器として利用し,相互補完させることで性能の向上を図る.実験では,提案手法を 3 種類の画像分類問題に適用した結果を示す.これらの実験から,提案手法の学習および検証用画像に対する分類精度が,従来の GIN-IC と比較して改善されることが確認できた.

When we evaluate the search performance of an evolutionary computation (EC) technique, we usually apply it to typical benchmark functions and evaluate its performance in comparison to other techniques. In experiments on limited benchmark functions, it can be diffcult to understand the features of each technique. In this paper, the search spaces that emphasize the performance difference of EC techniques are evolved by Cartesian genetic programming (CGP). We focus on a real-coded genetic algorithm (RCGA), which is a type of genetic algorithm that has a real-valued vector as a chromosome. The performance difference of two RCGAs is assumed to be a objective function of CGP, and the search space that increases the performance difference is evolved. In particular, we generate search spaces using the performance difference of real-coded crossovers or generation alternation models. As a result of our experiments, the search spaces that exhibit the largest performance difference of two RCGAs are generated for all the combinations. In addition, we extend the objective functions to two of the performance differences and the number of active nodes in CGP and attempt to generate multiple search spaces with an evolution using a multiobjective evolutionary algorithm. We then observe which types of elements expand the performance difference.

In recent years, because of the development of computers, it has been possible to analyze, that is to say the data mining. Due to this, there are a lot of studies about stock market prediction using past stock data. Almost all these methods, however, use only predicting brand information. There are a lot of factors of the price, but it is possible to think that other brands affect the price. In this paper, we propose a prediction method that uses not only the predicting brand information but also other brands information. We show the effectiveness of our method through the experiment of stock market prediction. © 2010 SICE.

The authors propose a three-layered network structure to detect abnormal objects in environments where surveillance cameras, security robots, and other image devices are employed for routine observations. By referring to the input patterns obtained from the environment, the network is structured to memorize the normal states of environments by constantly updating the connection weights in the network. As a result of learning, the network detects abnormal objects in input images. We conducted experiments in an office and in a corridor to verify the effectiveness of the proposed network for anomaly detection. © 2010 SICE.

Recently, the automation of picking work has advanced in the factory for the reduction of labor costs. When picking work is automatic, it is very important to estimate posture of target components. Therefore, an algorithm to automatically estimate posture is expected. We propose an efficient method to improve Differential Evolution (DE). Then, we applied the proposal method and DE to experiment to estimate posture of each four three-dimensional (3-D) objects by using the 3-D CAD data. The experimental results show that the proposal method achieves high success rate and efficiency than DE. © 2010 SICE.

In recent years, many researchers addressed multi-agent system. Multi-agent system is the system consisted of multiple robots that have only limited capability. Robocup simulated soccer is proposed as a test bed of multi-agent system. It has a subtask called keepaway soccer. Automatic construction of the strategy of multi-agent system is required because it is difficult. Therefore, we purpose construction of the strategy for multi-agent system by graph structured program evolution (GRAPE) in keepaway soccer. GRAPE is the method of construction of graph-structured programs automatically. © 2010 SICE.

The technology of three-dimensional (3-D) scene reconstruction is expected as applications of simulation, navigation, walk-through and so on. Besides, Segway®, a two-wheeled self-balancing electric vehicle, is being used in public facilities (e.g. airports, shopping centers, event sites). Therefore, it is beneficial to study applications using Segway. In this paper, we propose a method of 3-D scene reconstruction using Segway equipped with a stereo camera and a laser range finder (LRF). Our method estimates the location of Segway by a LRF and reconstruct 3-D environments by integrating 3-D shape information calculated by stereo vision. As the results of the experiments, the rough 3-D scene reconstruction results were obtained. © 2010 SICE.

In recently years, foreign exchange trading becomes active. Thus, financial products named foreign exchange(FX) was generated and many FX companies were established. Also the number of investors has increased because of the spread of transactions on the Internet. However, it is difficult to gain profits by FX, and many investors learn by mistakes to achieve an efficient transaction strategy. We propose a method to gain a efficient transaction strategy using Genetic Algorithm(GA) considering price trends. Then, this strategy is applied to foreign exchange transaction. The experiment using this strategy shows the effectivity. © 2010 SICE.

Automatic construction method for image classification algorithms have been required. Genetic image Network for Image Classification (GIN-IC) is one of the methods that construct image classification algorithms automatically, and its effectiveness has already been proven. In our study, we try to improve the performance of GIN-IC with Ad-aBoost algorithm using GIN-IC as weak classifiers to complement with each other. We apply our proposed method to three types of image classification problems, and show the results in tins paper. In our method, discrimination rates for training images and test images improved in the experiments compared with the previous method GIN-IC.

Automatic construction methods for image processing proposed till date approximate adequate image transformation from original images to their target images using a combination of several known image processing filters by evolutionary computation techniques. Genetic Image Network (GIN) is a recent automatic construction method for image transformation. The representation of GIN is a network structure. In this paper, we propose a method of automatic construction of image classifiers based on GIN, designated as Genetic Image Network for Image Classification (GIN-IC). The representation of GIN-IC is a feed-forward network structure. GIN-IC is composed of image transformation nodes, feature extraction nodes, and arithmetic operation nodes. GIN-IC transforms original images to easier-to-classify images using image transformation nodes, and selects adequate image features using feature extraction nodes. We apply GIN-IC to test problems involving multi-class categorization of texture images and two-class categorization of pedestrian and non-pedestrian images. Experimental results show that the use of image transformation nodes is effective for image classification problems.

Using well-established techniques of Genetic Programming (GP), we automatically optimize image feature filters over several inputs and within transformation images, improving the Automatic Construction of Tree-Structural Image Transformation (ACTIT) system. Our objective is to also produce optimal solutions in substantially less computation time than require for generating features of ACTIT. We improved the algorithm feature filters in the process through GP, which are expressed by trees in Automatic Construction of Tree-Structural Image Transformation, to reduce computation time. Through our experimentation, we show that our new approach is accurate and requires less computation time by maintaining the feature images in conjunction with the original images. © 2010 IEEE.

When we evaluate the search performance of an evolutionary computation (EC) technique, we usually apply it to typical benchmark functions and evaluate its performance in comparison to other techniques. In experiments on limited benchmark functions, it can be difficult to understand the features of each technique. In this paper, the search spaces that emphasize the performance difference of EC techniques are evolved by Cartesian genetic programming. We focus on a real-coded genetic algorithm, which is a type of genetic algorithm that has a real-valued vector as a chromosome. In particular, we generate search spaces using the performance difference of real-coded crossovers. In the experiments, we evolve the search spaces using the combination of three types of real-coded crossovers. As a result of our experiments, the search spaces that exhibit the largest performance difference of two crossovers are generated for all the combinations.

Image processing and recognition technologies are becoming increasingly important. Automatic construction methods for image transformation algorithms proposed to date approximate adequate image transformation from original images to their target images using a combination of several known image processing filters by evolutionary computation techniques. In this paper, we introduce the adaptive image processing filters that process according to the features of an input image. The processing of the adaptive filters is decided based on the local features of an input image. We implement them to feed-forward genetic image network (FFGIN) that is one of the automatic construction methods for image transformations. Then we apply our method to the problems of segmentation of organs and tissues in medical images. Experimental results show that our method constructs the effective segmentation algorithms that extract multiple regions respectively.

本論文では，複数の平面によって構成された物体を仮定し，照明効果による色度遷移についてのモデルを用いて，照明効果の影響による疑似輪郭線の発生を抑制した画像の領域分割手法を提案する．ひとつの表面に対応する画素群の色空間中での分布を光源との位置関係や Shadow，Shade など照明効果の影響によるものとみなし，隣接する領域が同一の表面に対応しているのか，それぞれが異なる表面に対応しているのか判定を行う．また，被写体表面を平面と想定できる環境を設定し，輪郭線の特性に基づいた領域の隣接関係についての判定を行うことで，隣接関係をもつ領域群から適切な統合先の選択を行う．照明効果から生じた疑似輪郭線によって過分割された領域を色度遷移と輪郭線の特性の情報をもとに適切に統合することで，物体表面色に基づいた画像領域分割を実現する．In this paper, we propose a method to segment a color image into regions using color transition model and plane approximation. We make the model of the change of chromaticity by the change of illumination condition. A color image is divided into several regions by the clustering results of color space. And, using the color transition model and plane approximation, all regions in the adjacency is determined whether or not they can be integrated. Integration of regions based on the result of determination, the whole image is divided into the regions which correspond to surfaces of object.

筆者らは先に，数種類の画像処理フィルタの適用順序および総数を遺伝的プログラミング （GP; Genetic Programming） によって最適化することで目的の画像処理を近似する方式を提案した．これを木構造状画像変換自動生成システム （ACTIT;Automatic Construction of Tree-structural Image Transformation） と呼ぶ．しかし，大量・多様の実画像を学習させた場合，すべての画像に対して有効な，1 つの画像処理を構築することは難しい．そこで，本報告では，アンサンブル学習法の一種である AdaBoost （Adaptive Boosting） を ACTIT に応用し，複数の木構造状画像変換を組み合わせることによって有効な処理を構築するシステム，ACTIT-Boost を提案する．本手法は AdaBoost の理論に基づいているため，十分な数の木構造状画像変換を構築することで，学習した画像に対しては目標となる画像に正しく変換が行なえる処理となることが期待される．We have already proposed the system which automatically constructs image processing with Genetic Programming (GP). It was named Automatic Construction of Tree-structural Image Transformation (ACTIT). However, it is difficult to construct an accurate image processing for all training image sets in case there are many and various images. It is necessary to combine many subroutines of image processing. In this paper, we propose ACTIT-Boost which automatically constructs an accurate image processing by employing Adaptive Boosting (AdaBoost) to ACTIT. If there are enough number of weak classifier, AdaBoost has been proved to be extremely successful in producing accurate classifiers. Therefore, ACTIT-Boost constructs a perfect image processing for training image sets.

本論文では，複数の平面によって構成された物体を仮定し，照明効果による色度遷移についてのモデルを用いて，照明効果の影響による疑似輪郭線の発生を抑制した画像の領域分割手法を提案する．ひとつの表面に対応する画素群の色空間中での分布を光源との位置関係や Shadow，Shade など照明効果の影響によるものとみなし，隣接する領域が同一の表面に対応しているのか，それぞれが異なる表面に対応しているのか判定を行う．また，被写体表面を平面と想定できる環境を設定し，輪郭線の特性に基づいた領域の隣接関係についての判定を行うことで，隣接関係をもつ領域群から適切な統合先の選択を行う．照明効果から生じた疑似輪郭線によって過分割された領域を色度遷移と輪郭線の特性の情報をもとに適切に統合することで，物体表面色に基づいた画像領域分割を実現する．In this paper, we propose a method to segment a color image into regions using color transition model and plane approximation. We make the model of the change of chromaticity by the change of illumination condition. A color image is divided into several regions by the clustering results of color space. And, using the color transition model and plane approximation, all regions in the adjacency is determined whether or not they can be integrated. Integration of regions based on the result of determination, the whole image is divided into the regions which correspond to surfaces of object.

筆者らは先に，数種類の画像処理フィルタの適用順序および総数を遺伝的プログラミング （GP; Genetic Programming） によって最適化することで目的の画像処理を近似する方式を提案した．これを木構造状画像変換自動生成システム （ACTIT;Automatic Construction of Tree-structural Image Transformation） と呼ぶ．しかし，大量・多様の実画像を学習させた場合，すべての画像に対して有効な，1 つの画像処理を構築することは難しい．そこで，本報告では，アンサンブル学習法の一種である AdaBoost （Adaptive Boosting） を ACTIT に応用し，複数の木構造状画像変換を組み合わせることによって有効な処理を構築するシステム，ACTIT-Boost を提案する．本手法は AdaBoost の理論に基づいているため，十分な数の木構造状画像変換を構築することで，学習した画像に対しては目標となる画像に正しく変換が行なえる処理となることが期待される．We have already proposed the system which automatically constructs image processing with Genetic Programming (GP). It was named Automatic Construction of Tree-structural Image Transformation (ACTIT). However, it is difficult to construct an accurate image processing for all training image sets in case there are many and various images. It is necessary to combine many subroutines of image processing. In this paper, we propose ACTIT-Boost which automatically constructs an accurate image processing by employing Adaptive Boosting (AdaBoost) to ACTIT. If there are enough number of weak classifier, AdaBoost has been proved to be extremely successful in producing accurate classifiers. Therefore, ACTIT-Boost constructs a perfect image processing for training image sets.

本研究の目的は,人間が持つカテゴリカル色知覚と色の恒常性の機能を備えたモデルを獲得し,これを用いて未知の照明光下で撮影された画像中の物体色の色認識を行うことである.さまざまな照明条件で撮影された画像中の色認識はコンピュータビジョンやマシンビジョンに有効であり非常に重要であるが,色恒常性を解決する手法はいまだに確立されておらず,特に,人間の色覚特性を調べた心理物理実験の結果を利用した手法は提案されていない.そこで我々は,さまざまな照明条件下で行われたカテゴリカルカラーネーミング実験の結果をニューラルネットワークで学習することで色恒常性も考慮したカテゴリカル色知覚モデルを獲得し,このモデルを画像中の各画素の色認識に適用してモデルが人間と同じように色認識をできていることを示す.

筆者らはこれまでに,人間がもつカテゴリカル色知覚と色の恒常性の機能を備えたモデルを,ニューラルネットワークを用いて構築してきた.これらの研究では,心理物理実験結果を用いることで色覚正常者のカテゴリカル色知覚モデルを構築した,本研究では,色覚異常者の色知覚モデルを構築することを目的としている.色覚異常者の色知覚モデルを構築し,そのメカニズムを解析することで,人間の色覚メカニズムを考祭することができると考えている.本報告では色覚異常者の心理物理実験結果をもとに,構造が任意のニューラルネットワークをGAによって最適化するFFFCNを用いて,モデルの構築を試みる.

The purpose of this study is to get the model that can operate similarly to human categorical color perception and color constancy. It is related to a categorical color perception system for automatically judging a categorical color. And it is a technology for correctly judging under various environments. It is important for security camera in particular. To obtain the model of categorical color perception and color constancy, the relationship, between the chromaticity and the categorical color perception of colored chips under different illuminations, is trained by using an artificial neural network. The categorical color perception is the product of a categorical color naming experiment. As a result, we propose a categorical color perception system which can correctly judge a categorical color under various environments. In addition, the trained network determined a color-name of objects from images. Experimental results show that the obtained neural network had the color recognition similar to human vision system.

本研究の目的は,人間がもつカテゴリカル色知覚と色の恒常性の機能を備えたモデルを獲得し,これを用いて未知の照明光下で撮影された画像中の物体色の色認識を行うことである.さまざまな照明条件で撮影された画像中の色認識はコンピュータビジョンやマシンビジョンに有効であり非常に重要であるが,色恒常性を解決する手法は未だに確立されておらず,特に,人間の色覚特性を調べた心理物理実験の結果を利用した手法は提案されていない.そこで我々は,様々な照明条件下で行われたカテゴリカルカラーネーミング実験の結果をニューラルネットワークで学習することで色恒常性も考慮したカテゴリカル色知覚モデルを獲得し,このモデルを画像中の各画素の色認識に適用してモデルが人間と同じように色認識をできていることを示す.

We developed a model that can operate similarly to human categorical color perception. The color of an object is not exclusively determined by the reflection spectrum from the surface of the object but is greatly affected by the ambient environmental conditions and depends upon color constancy. The mechanism of color constancy, however, is not explained in detail so acquiring the cognition of the categorical color name of objects under different illuminations is difficult. To that end, the relationship between the chromaticity and the categorical color perception of colored chips under different illuminations is the product of a categorical color-naming experiment was learned by using a neural network. The results showed that the obtained neural network has similar characteristics to those of human vision system.

We developed a model that can operate similarly to human categorical color perception. The color of an object is not exclusively determined by the reflection spectrum from the surface of the object but is greatly affected by the ambient environmental conditions and depends upon color constancy. The mechanism of color constancy, however, is not explained in detail so acquiring the cognition of the categorical color name of objects under different illuminations is difficult. To that end, the relationship between the chromaticity and the categorical color perception of colored chips under diff...