Computer Vision In Robotics And Industrial Applications

Dominik Sankowski, Jacek Nowakowski

• Publisher: WSPC
• Year: 2014
• ISBN:
  9789814583732
• Topic: Biological Sciences
• Subtopic: Science General
• Index: Biological Sciences

The book presents a collection of practical applications of image processing and analysis. Different vision systems are more often used among others in the automotive industry, pharmacy, military and police equipment, automated production and measurement systems. In each of these fields of technology, digital image processing and analysis module is a critical part of the process of building this type of system.

The majority of books in the market deal with theoretical issues. However, this unique publication specially highlights industrial applications, especially industrial measurement applications. Along with its wide spectrum of image processing and analysis applications, this book is an interesting reference for both students and professionals.

Contents:

• Theoretical Introduction to Image Reconstruction and Processing:
  • Data Set Preparation for k -NN Classifier Using the Measure of Representativeness (Marcin Raniszewski)
  • Segmentation Methods in the Selected Industrial Computer Vision Application (Anna Fabijanska and Dominik Sankowski)
  • Line Fractional-Order Difference/Sum, Its Properties and an Application in Image Processing (Piotr Ostalczyk)
• Computer Vision in Robotics:
  • Management Software for Distributed Mobile Robot System (Maciej Łaski, Sylwester Błaszczyk, Piotr Duch, Rafał Jachowicz, Adam Wulkiewicz, Dominik Sankowski and Piotr Ostalczyk)
  • Advanced Vision Systems in Detection and Analysis of Characteristic Features of Objects (Adam Wulkiewicz, Rafał Jachowicz, Sylwester Błaszczyk, Piotr Duch, Maciej Łaski, Dominik Sankowski and Piotr Ostalczyk)
  • Pattern Recognition Algorithms for the Navigation of Mobile Platform (Rafał Jachowicz, Sylwester Błaszczyk, Piotr Duch, Maciej Łaski, Adam Wulkiewicz, Dominik Sankowski and Piotr Ostalczyk)
  • Partial Fractional-Order Difference in the Edge Detection (Piotr Duch, Rafał Jachowicz, Sylwester Błaszczyk, Maciej Łaski, Adam Wulkiewicz, Piotr Ostalczyk and Dominik Sankowski)
  • Application of Fractional-Order Derivative for Edge Detection in Mobile Robot System (Sylwester Błaszczyk, Rafał Jachowicz, Piotr Duch, Maciej Łaski, Adam Wulkiewicz, Piotr Ostalczyk and Dominik Sankowski)
  • Vision Based Human-Machine Interfaces: Visem Recognition (Krzysztof Ślot, Agnieszka Owczarek and Maria Janczyk)
• Industrial Applications of Computer Vision in Process Tomography, Material Science and Temperature Control:
  • Hybrid Boundary Element Method Applied for Diffusion Tomography Problems (Jan Sikora, Maciej Pańczyk and Paweł Wieleba)
  • Two-phase Gas-Liquid Flow Structures and Phase Distribution Determination Based on 3D Electrical Capacitance Tomography Visualization (Robert Banasiak, Radosław Wajman, Tomasz Jaworski, Paweł Fiderek, Jacek Nowakowski and Henryk Fidos)
  • Tomographic Visualization of Dynamic Industrial Solid Transporting and Storage Systems (Zbigniew Chaniecki, Krzysztof Grudzień and Andrzej Romanowski)
  • Tomography Data Processing for Multiphase Industrial Process Monitoring (Krzysztof Grudzień, Zbigniew Chaniecki, Andrzej Romanowski, Jacek Nowakowski and Dominik Sankowski)
  • Dedicated 3D Image Processing Methods for the Analysis of X-Ray Tomography Data: Case Study of Materials Science (Laurent Babout and Marcin Janaszewski)
  • Selected Algorithms of Quantitative Image Analysis for Measurements of Properties Characterizing Interfacial Interactions at High Temperatures (Krzysztof Strzecha, Anna Fabijańska, Tomasz Koszmider and Dominik Sankowski)
  • Theoretical Introduction to Image Reconstruction for Capacitance Process Tomography (Radosław Wajman, Krzysztof Grudzien, Robert Banasiak, Andrzej Romanowski, Zbigniew Chaniecki and Dominik Sankowski)
  • Infra-Red Thermovision in Surface Temperature Control System (Jacek Kucharski, Tomasz Jaworski, Andrzej Frączyk and Piotr Urbanek)
• Medical and Other Applications of Computer Vision:
  • The Computer Evaluation of Surface Color Changes in Cultivated Plants Influence by Different Environmental Factors (Joanna Sekulska-Nalewajko and Jarosław Gocławski)
  • Various Approaches to Processing and Analysis of Images Obtained from Immunoenzymatic Visualization of Secretory Activity with ELISPOT Method (Wojciech Bieniecki and Szymon Grabowski)
  • Image Processing and Analysis Algorithms for Assessment and Diagnosis of Brain Diseases (Anna Fabijanska and Tomasz Węglinski)
  • Computer Systems for Studying Dynamic Properties of Materials at High Temperatures (Marcin Bąkała, Rafał Wojciechowski and Dominik Sankowski)

Readership: Researchers, professionals and academics in image analysis, machine perception/computer vision, software engineering and fuzzy logic.

Part 1

Theoretical Introduction to Image Reconstruction and Processing

CHAPTER 1

DATA SET PREPARATION FOR K-NN CLASSIFIER USING THE MEASURE OF REPRESENTATIVENESS

Marcin Raniszewski

Institute of Applied Computer Science
Lodz University of Technology
90-924 Łódź, ul. Stefanowskiego 18/22
[email protected]

In data classification a decision is made on the basis of information of the data transfer (described by a set of features). Proper and rapid classification depends on the preparation of the data set, as well as the selection of classification algorithms. The k nearest neighbors algorithm (k-NN) is one of the most popular classification algorithms: simple to implement and intuitive. The k-NN has high classification accuracy in a wide range of applications. But in the standard k-NN a complete data set is used during the classification. This causes problems with large data sets: high memory requirements and the speed of classification decrease. To eliminate these problems reduction and editing algorithms can be used. Reduction algorithms remove a significant part of the data. The remaining samples must provide an acceptable level of classification accuracy. Editing algorithms filter a data set, removing noise and redundant data.

This chapter presents reduction and editing algorithms, both using a measure of representativeness. Each proper sample of a data set is a part of its class distribution. A measure of representativeness expresses the level of this information retained by a sample.

The presented algorithms were tested on seven real data sets well known in the literature. The results are promising in comparison with the results of popular reduction and editing algorithms.

1.1 Introduction

The Nearest Neighbor Rule (NN), the particular case of the k Nearest Neighbor Rule for k = 1, is still in use in many applications of Pattern Recognition (Duda, Hart and Stork 2001; Theodoridis and Koutroumbas 2006). This method of classification is intuitive, simple and effective: a new sample is classified to a class of its nearest neighbor (the closest sample) from a training set. In (Duda, Hart and Stork 2001) the authors proved that an NN classification error is never beyond a double classification error of the Bayesian classifier for sufficiently large training sets. NN has also other advantages: the lack of training phase and faster classification as opposed to k-NN with k > 1.

On the other hand, NN (like standard k-NN) uses a complete training set to classify new samples. It causes computational loads and increases memory requirements for large training sets (e.g. in image analysis). It is a serious problem in applications with large reference sets, where the speed of classification is crucial.

A well-known solution of the problem is the reduction of a reference set: the majority of samples are removed and only the most important samples are preserved (Wilson and Martinez 2000). Obviously, the classification accuracy of NN operating on the reduced reference set should be acceptable. After reduction the classification accuracy can even increase for certain data sets as a consequence of removal of a noise.

There are many popular and widely used reduction methods. The newer methods use heuristics such as Monte Carlo, Random Mutation Hill Climbing, Tabu Search or Genetic Algorithms (Skalak 1994, Kuncheva and Bezdek 1998, Cerveron and Ferri 2001) and despite good results they provide, all of them are random and have several parameters. The aim of presented research was to construct a non-random reduction algorithm providing similarly good results as heuristic procedures, with fewer parameters and a much shorter training phase than Tabu Search or genetic methods. The proposed reduction algorithm uses the idea of the most representative samples. They are selected according to the measure of representativeness.

The NN procedure is very vulnerable to a noise. All new samples for which the noise samples are nearest neighbors are misclassified. The editing algorithm tries to solve this problem: noise, mislabelled and atypical samples are recognized by the algorithm and removed from the reference set. The results of proposed reduction method based on the most representative samples was so promising that the idea of representativeness was also used to create an editing method.

Proposed in this paper reduction and editing algorithms were tested on seven real data sets and compared with the results of other well-known methods.

1.2 Well-known reduction techniques

Probably the most popular reduction algorithm is Condensed Nearest Neighbor Rule (CNN) described by Hart (Hart 1968). The algorithm produces a consistent reference set, which means that it correctly classifies (using NN) all samples from a training set. Hart called it a consistency criterion. This criterion was used for a long time in other reduction methods as a reliable stop condition in creating a reduced reference set.

Gates proposed Reduced Nearest Neighbor Rule (RNN) (Gates 1972). This method uses the result of the CNN algorithm and returns a consistent subset of the CNN reduced set.

Gowda and Krishna published a number of articles about different uses of the concept of Mutual Nearest Neighborhood (MNN) and Mutual Neighborhood Value (MNV). One of their papers concerns a reference set reduction method (GK) (Gowda and Krishna 1979). Gowda and Krishna’s method also results in a consistent reduced set.

Dasarathy used the concept of Nearest Unlike Neighbor (NUN): the sample x is NUN(y) if it is from a different class than sample y and is y’s nearest neighbor (Dasarathy 1994). Dasarathy believed his procedure (MCS) results in minimal consistent reduced sets. Kuncheva and Bezdek presented in (Kuncheva and Bezdek 1998) a counter-example.

Skalak proposed two heuristics: MC1 and RMHC-P (Skalak 1994). The former is based on Monte Carlo heuristic, while the latter on Random Mutation Hill Climbing. In RMHC-P initially m samples are selected to a reduced set and n replacements (mutations) are made: random samples from a reduced set are replaced with random samples (which is not actually in the reduced set) from a training set. If the replacement increases k-NN classification accuracy counted for all the samples from the training set with the actual reduced set as a reference set, then the replacement is accepted, otherwise rejected. After all n replacements, the procedure finishes. The reduced set contains m samples and is inconsistent.

Kuncheva described heuristics based on genetic algorithms (GA) (Kunche...

Table of contents

• Cover Page
• Title Page
• Copyright Page
• Contents
• Preface
• Part 1 Theoretical Introduction to Image Reconstruction and Processing
• Part 2 Computer Vision in Robotics
• Part 3 Industrial Applications of Computer Vision in Process Tomography, Material Science and Temperature Control
• Part 4 Medical and Other Applications of Computer Vision
• Index