Back to search

Democritus University of Thrace (DUTH)

Ανάπτυξη συστημάτων νευρωνικών πολυταξινομητών για αναγνώριση αντικειμένων και εφαρμογή τους στη ρομποτική όραση

Abstract

dc:description

This dissertation is concerned with the development of an automatic handling and assembling system combining the use of a fixed base manipulator with an image processing and recognition system. To this end, the recognition of objects in the working environment of the manipulator is of paramount importance and guides the development of the work in this thesis. The recognition system uses a two-dimensional image representation of the objects and extracts appropriate object features, while in a second level a classification scheme is used in order to recognize the objects. New object feature vectors are proposed, which are based on the polygon approximation of the object shape and a new neural multi-classifier scheme is proposed aiming at the enhancement of the recognition accuracy. Basic components of the recognition system are image preprocessing, feature vector extraction and object classification. During the image preprocessing phase the various objects are initially detected and in the sequel transformed so that useful parameters (centres of holes, area, center of area, perimeter, rotation angle) can be calculated. These parameters are in the sequel used in order to derive useful feature vectors, or are used as supplemental information for the guidance of the manipulator. Feature extraction is based on descriptors that use one dimensional representation of the surrounding line of a 2-dimensional figure, aiming at the production of reliable features with low computational cost. Descriptors are based on geometrical features, on statistical moments and on polygon approximation of object shapes. This way, three independent feature vectors are formed, which are invariant to the position, the size and the orientation of image objects. The first vector is composed of three complementary geometrical features based on the area of the figure its perimeter and the distances of the contour points to the center of area. The second vector is formed by using scaled normalized central moments, which arrive from statistical moments and do not suffer from the exponential growth presented by high order moments. The last vector is composed of features produced by a stepwise polygon approximation method, which was derived in this dissertation aiming at the real time extraction of reliable features. Stepwise polygon approximation offers a satisfactory solution to the problem of optimally determining a polygon with predefined number of edges able to approximate a given shape contour. The proposed methodology is particularly useful in situations that time of computation is critical. It uses a technique of repeating approximations with successively reducing input data points. Following this procedure two feature vectors are extracted, which use the angles of the approximating polygon and the successive approximation errors. These errors are defined to be the ratio of the initial contour length to the length of an edge determined by two successive vertexes of the polygon. The two features are invariant to space, magnitude and orientation of the shape they describe. Therefore, are mighty features and can be used either independently or in combination with other features in 2-dimensional shape recognition applications. In the last stage of the recognition procedure, a multi-classifier scheme is proposed, which is based on neural classifiers and comprises two levels. In the first level, neural classifiers are trained to perform independently of each other, based on different feature vectors. In the second level the final classification is performed, combining the outputs of the first level classifiers and using a simplified layout of small neural networks. More specifically, a low order neural network is assigned to all similar output classes of the independent first level classifiers and is trained to perform the final classification. This way, the ability of each first level classifier to differentiate its results based on the appearance of non-trained data is better taken into consideration in the final judgment. The proposed architecture is of low complexity presenting at the same time high recognition rates. Moreover, it requires small number of training cycles, making thus the recognition scheme appropriate for demanding classification applications, where in a changing environment the system needs retraining during its operation. The performance of the proposed multi-classifier architecture is tested and compared to other existing techniques based on an adequate number of digitally modified patterns. Finally, the proposed methodologies are tested by building a robotic vision application experiment using the existing industrial manipulator of the laboratory of automatic control systems and robotics of Democritus University of Thrace.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2008

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Mitzias, Dimitrios
  • Μήτζιας, Δημήτριος

Subjects

dc:subject × 8

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/20960

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Mitzias, Dimitrios; Μήτζιας, Δημήτριος. Ανάπτυξη συστημάτων νευρωνικών πολυταξινομητών για αναγνώριση αντικειμένων και εφαρμογή τους στη ρομποτική όραση. Democritus University of Thrace (DUTH), 2008. http://hdl.handle.net/10442/hedi/20960