Image Recognition doesn't work

Anonymous
2012-04-18
2012-12-24

  • Anonymous
    2012-04-18

    **i have 3 sets of images on which i have trained neuroph. But when i try to recognize random images that is completely not matching my sample data set, it gets an accuracy nearing to 98%. How can this happen?

    for example my training data set is a sample set of car images but when i take the picture of a computer mice it recognises it as one of the cars with 98% accuracy.

    Can some one help me why it is happening.**

     
  • Zoran Sevarac
    Zoran Sevarac
    2012-04-19

    Use so called junk images, or in other words specify images that should not be recognized
    Ususally these are all red, all green and all blue images in order to avoid unwanred recognition in case of high activation levels on all inputs.
    Since neuroph is using rgb data as inputs, all white image woul dput high activation on all neurons and it would result as recognition, In order to avoid this we put these in  junk images. You can specify these images using wizard in Neuroph Studio.

    Zoran

     

  • Anonymous
    2012-04-20

    Thanks zoran.

    Also i would like to know what is the correct behaviour and what would be the result in percentages , if i add an image as junk and give that same image to neuroph for recognition.

    will i get 0 as the result???

    Thanks in Advance

     
  • Zoran Sevarac
    Zoran Sevarac
    2012-04-22

    Yes, usually something close to zero ( or far enough from 1)

     

  • Anonymous
    2012-04-23

    ok.. thanks.

    Not sure ..but in most cases for me it doesn't play well . added to junk still get a match from the trained set and values are not so close enough to zero.

    Any suggestions on how can i improve the results zoran.

     
  • Zoran Sevarac
    Zoran Sevarac
    2012-04-25

    Try different number of hidden neurons
    Try different learning rate, momentum
    If that doesent work try adding more layers of hidden neurons…