Co-operative Co-evolution of Image Feature Extractors and Object Detection Algorithms

Created by W.Langdon from gp-bibliography.bib Revision:1.7630

@PhdThesis{Roberts:thesis,

• author = "Mark E. Roberts",
• title = "Co-operative Co-evolution of Image Feature Extractors and Object Detection Algorithms",
• school = "School of Computer Science, University of Birmingham",
• year = "2006",
• address = "UK",
• month = feb,
• keywords = "genetic algorithms, genetic programming",
• URL = "https://britishmachinevisionassociation.github.io/information/theses",
• URL = "http://www.bmva.org/theses/2006/2006-roberts.pdf",
• size = "185 pages",
• abstract = "Machine learning and pattern recognition problems are commonly decomposed into twostages: feature extraction and classification. The feature extraction stage is often manually designed using knowledge of the problem domain. This results in biases and inflexibility especially in areas such as image analysis where the raw data is large and complex.

  This thesis presents a novel framework for simultaneously co-evolving both the feature extraction stage and the classification stage. A representation is defined which allows the evolution of statistical image features. The classifiers, evolved using genetic programming, take the form of programs which manipulate the extracted features. The architecture ofthe coevolutionary system is defined along with the underlying collaboration and credit assignment mechanisms, which allow the two stages to implicitly co-operate.

  The system is evaluated on synthetic datasets which exhibit noise and variation inscale and orientation, and on complex natural images. Two different training methodsare used. In the first, traditional method, every pixel in the image is used for training.Although successful, this method suffers from high computation costs and problems with scaling to more complex problems. In the second method, a multi-stage sampling scheme is used which requires substantially less computation. This method demonstrates similar or better performance than the previous method and is more effective at solving complex problems. In both methods, the co-evolutionary approach is shown to produce solutions that are significantly better than when using a fixed set of features.

  The success of the approach developed in this work shows that the coevolution of the two-stage learning processes is possible in this domain, a finding which may have implications in other areas of artificial intelligence and pattern recognition.",

}

Genetic Programming entries for Mark E Roberts

Citations