Paper Info
Title
Performing Image Classification with a Frequency-based Information Retrieval Schema for ImageCLEF 2006.
Abstract
This article describes the participation of the University and Hospitals of Geneva at the ImageCLEF 2006 image classication tasks (medical and non{medical). The tech- niques applied are based on classical tf/idf weightings of visual features as used in the GIFT (GNU Image Finding Tool) image retrieval engine. Based on the training data, features that appear in images of the same class are weighted higher than features appearing across many classes. These feature weights are added to the classical ft/idf weights, making it a mixture of weightings. Several weightings and learning approaches are applied as well as several quantisations of the features space with respect to grey levels. A surprisingly small number of grey levels leads to best results. Learning can improve the results only slightly and does not obtain as good results as classical image classication approaches. A combination of several classiers leads to best nal results, showing that the applied schemes have independent results. For future work it seems important to study in more detail the important features and feature groups as they are not independent in the GIFT system. Pre{treating of the images (background removal) or allowing for more variation of the images with respect to object size and position might be other approaches to further improve results.
Year
Venue
Keywords
2006
CLEF (Working Notes)
frequency{based weights,image classication,image retrieval,feature space,information retrieval
Field
DocType
Citations 
Small number,Training set,Information retrieval,Computer science,Image retrieval,Contextual image classification,Schema (psychology),Visual Word
Conference
3
PageRank 
References 
Authors
0.45
6
3
Name
Order
Citations
PageRank
henning muller118324.26
Tobias Gass214311.91
Antoine Geissbuhler381549.75