Home > About Us > Technology > Technical Report > Technical Report No.22 2013

Automatic Image Annotation with Probabilistic Generative Models

We are developing automatic image annotation technology that assigns labels to images for image retrieval. In this report, we propose two algorithms that use probabilistic generative models that are effective even for a small number of training images. The first algorithm reduces overfitting for labels associated with a small number of images by maximizing the cross entropy of the models for those labels, thereby achieving the highest performance as an algorithm using generative models. The other algorithm achieved fast training and testing by using Random Forest classifiers for the estimation of probabilities.

Author

  • Noriji Kato
    Communication Technology Laboratory, Research & Technology Group
  • Motofumi Fukui
    Communication Technology Laboratory, Research & Technology Group
  • Yukihiro Tsuboshita
    Communication Technology Laboratory, Research & Technology Group
  • Ryota Ozaki
    Communication Technology Laboratory, Research & Technology Group