Krzysztof Krawiec


Home

Research:

edit SideBar

This thesis concentrates in the fields of computer vision, image understanding and machine learning. We present a novel approach for learning visual concepts from raw image data. Our method is able to learn new concepts and, in result, acquire knowledge that can be than used in standard computer vision tasks such as pattern recognition, object identification or object tracking. The acquired knowledge is encoded in a form of individuals/learners which are able to process visual information. The originality of this approach lies also in the fact that individuals do not process raw image data directly. Instead, they operate on a set of attributed visual primitives that are acquired from images in the preliminary stage of the processing. The proposed approach is very general, because it uses the learning principle. For the purpose of learning we use genetic programming. Each individual/learner represents a procedure in a form of a tree of operations on sets of visual primitives. The images from the training set in our approach are not labeled in any way and no information is assigned with them. It means that the knowledge is acquired in totally unsupervised way. The only feedback information that states which individual better represents the target concept from the training set is the measure how the individual is able to reconstruct the input image. The rational standing behind this approach is following. We believe that a visual concept can be proved to be acquired and understood only if it can be reconstructed and successfully compared with original image from it was previously learned. In the thesis we present our methodology in details and describe its exemplary implementation that is based on the concept of segment as visual primitive. The method is verified on several sets of shapes of simple objects such as triangles, sections, Y letters and others. We also elaborate a methodology of cross-task knowledge sharing that is a step towards modularization of knowledge and provide results of massive experiments with cross-task knowledge sharing. Finally, the most important design decisions of software environment that was developed as a platform for computational experiments with our approach are described. A particular emphasis is put on computational performance issues of the developed software.

@MASTERSTHESIS { jaskowski06crosstask,
    AUTHOR = { Wojciech Ja\'skowski },
    TITLE = { Genetic Programming with Cross-task Knowledge Sharing for Learning of Visual Concepts },
    SCHOOL = { Poznan University of Technology },
    YEAR = { 2006 },
    ADDRESS = { Pozna\'n, Poland },
    MONTH = { jul },
    ABSTRACT = { This thesis concentrates in the fields of computer vision, image understanding and machine learning. We present a novel approach for learning visual concepts from raw image data. Our method is able to learn new concepts and, in result, acquire knowledge that can be than used in standard computer vision tasks such as pattern recognition, object identification or object tracking. The acquired knowledge is encoded in a form of individuals/learners which are able to process visual information. The originality of this approach lies also in the fact that individuals do not process raw image data directly. Instead, they operate on a set of attributed visual primitives that are acquired from images in the preliminary stage of the processing. The proposed approach is very general, because it uses the learning principle. For the purpose of learning we use genetic programming. Each individual/learner represents a procedure in a form of a tree of operations on sets of visual primitives. The images from the training set in our approach are not labeled in any way and no information is assigned with them. It means that the knowledge is acquired in totally unsupervised way. The only feedback information that states which individual better represents the target concept from the training set is the measure how the individual is able to reconstruct the input image. The rational standing behind this approach is following. We believe that a visual concept can be proved to be acquired and understood only if it can be reconstructed and successfully compared with original image from it was previously learned. In the thesis we present our methodology in details and describe its exemplary implementation that is based on the concept of segment as visual primitive. The method is verified on several sets of shapes of simple objects such as triangles, sections, Y letters and others. We also elaborate a methodology of cross-task knowledge sharing that is a step towards modularization of knowledge and provide results of massive experiments with cross-task knowledge sharing. Finally, the most important design decisions of software environment that was developed as a platform for computational experiments with our approach are described. A particular emphasis is put on computational performance issues of the developed software. },
    FILE = { jaskowski06crosstask.pdf:\\jaskowski06crosstask.pdf:PDF },
    URL = { http://www.cs.put.poznan.pl/wjaskowski/pub/papers/jaskowski06crosstask.pdf },
    KEYWORDS = { evolutionary computation, genetic programming, visual learning, knowledge sharing, knowledge reuse },
}


Powered by PmWiki