Classification

  1. Result

    1. Confusion Matrix

      2. Here below, the confusion matrix for all views and best network architecture (inputs, number of hidden nodes)

    2. Analysis

      3. Having tried several inputs, looked at the classification results and the network outputs, we preconise to use as inputs:
      • detail view angle
      • lms distance
      • detail energy
      Several numbers of hidden nodes have been tried. We see that the result is not very sensitive to that number and that -10 (10 time less weights/unknowns than patterns/equations) works well. To be convinced no over-training has occured, we may:
      • look at networks output
      • look at classification results for a split database (odd images for training, even for testing)

  2. Generalisabilty

    The network learns:

    The first term is generalisable because:

    • measures invariant to gain/contrast have been used.
    • net learned:
      • typical values for a clear detail - no detail
      • good interpolation
    A detail may not be found because:
    • the detail is occluded
    • the detail is not visible
    • the detail is visible but the search was not successful

    The processing scheme is robust to limited occlusion but problems arise for bad quality images because the probability of not finding a detail (learned on IRIS database) is not generalisable:

    • it is low on good images (IRIS)
    • it is high on bad quality images (Mepen95).

    To overcome this problem, we may:

    • learn on a more realistic database (not available)
    • Truncate probability to 0.5:
      • we 'vote' for a vehicle if we believe a detail has been found (appropriate measures)
      • but we don't 'vote' against if a detail has not been found because that detail could be there but not found).