Loading...

Active Learning in Image Retrieval

Fadaee, Mohsen | 2013

985 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 44107 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiei, Hamid Reza
  7. Abstract:
  8. Image retrieval, simply put, is the process of finding images in a predefined set , that are similar to an image specified by the user. In particular, the user inputs an image as query, and expects to see images similar to the query. Our purpose is to retrieve the images, by means of visual features, without any use of latent information such as tags and annotations.Afer the first round of retrieval, the answers can become more accurate, by means of user feedbacks. In this state, using active learning methods may be usefull. By using active data selection, we hope to achieve the answer faster. Learning based on manifold assumption, is another means which may be used in image retrieval. Exploiting this assumption and considering the structure of data, it is possible to use numerous unlabeled data. In this research our effort is to achieve an image retrieval algorithm based on manifold assumption the first challenge, which we confronted, was imbalanced training data. This occures when we wish to retrieve an image among a dataset with plenty of classes. Using the Gaussian pro-cesses framework, we formulated manifold based image retrieval. Therefore, we can estimate and remove the imbalance from the data. The results of our experiments shows a consistent improvement in all datasets.Another challenge, we focused on was selecting a batch of images for user feedback. In batch selection, user feedback is provided per images, and then the batch of images must to be selected with minimum relevant information overlap. To that end we used a method that selects from a set of mages, those that best represent the entire set. The experiments shows the improvement over the base method and k-means as another batch selection method
  9. Keywords:
  10. Active Learning ; Relevance Feedback

 Digital Object List

 Bookmark

...see more