Information

Abstract

In this thesis, we developed an image classification model with improving classification performance over a training phase. The model is using a pre-trained convolutional neuronal network (CNN) for feature extraction and a k-means algorithm for clustering. Performance optimization is realized by optimized weight factors for the extracted feature values. The optimization of the weight factors is calculated iteratively during a training phase. The measure of the weight factor adoption in a training step is related to the ground-truth dependent clustering contribution of the newly added image feature. We see as an advantage of our approach that the optimization requires no internal changes of the applied feature extraction and clustering algorithms, hence pre-trained models or closed-source implementations can be used. As a further advantage, we see the step-wise transparency of the performance development during the training phase for each newly added image as opposed to batch-based training for CNNs. This enables dynamic control of the training phase by the user. Another advantage is the small number of parameters to be optimized, which results in reduced processing time. A further advantage is the classification performance of our model that outperforms the reference model without feature weight optimization. In the course of our work, we developed a Python application that implements our model and provides a user-friendly interface. It allows easy set-up of test cases and provides graphics and tables for a comprehensive evaluation on process steps level. We consider this application as a starting point for future work.

Additional Files and Images

Additional images and videos

Additional files

Weblinks

No further information available.

BibTeX

@bachelorsthesis{Gruber2019,
  title =      "Extended Image Classification",
  author =     "Horst Gruber",
  year =       "2019",
  abstract =   "In this thesis, we developed an image classification model
               with improving classification performance over a training
               phase. The model is using a pre-trained convolutional
               neuronal network (CNN) for feature extraction and a k-means
               algorithm for clustering. Performance optimization is
               realized by optimized weight factors for the extracted
               feature values. The optimization of the weight factors is
               calculated iteratively during a training phase. The measure
               of the weight factor adoption in a training step is related
               to the ground-truth dependent clustering contribution of the
               newly added image feature. We see as an advantage of our
               approach that the optimization requires no internal changes
               of the applied feature extraction and clustering algorithms,
               hence pre-trained models or closed-source implementations
               can be used. As a further advantage, we see the step-wise
               transparency of the performance development during the
               training phase for each newly added image as opposed to
               batch-based training for CNNs. This enables dynamic control
               of the training phase by the user. Another advantage is the
               small number of parameters to be optimized, which results in
               reduced processing time. A further advantage is the
               classification performance of our model that outperforms the
               reference model without feature weight optimization. In the
               course of our work, we developed a Python application that
               implements our model and provides a user-friendly interface.
               It allows easy set-up of test cases and provides graphics
               and tables for a comprehensive evaluation on process steps
               level. We consider this application as a starting point for
               future work.",
  month =      aug,
  address =    "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria",
  school =     "Research Unit of Computer Graphics, Institute of Visual
               Computing and Human-Centered Technology, Faculty of
               Informatics, TU Wien ",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2019/Gruber2019/",
}