Probabilistic Normalization: An Approach to Normalizing Similarity Measures in Content based Image Retrieval

M. Arevalillo-Herráez, J. Domingo, and M. Zacarés (Spain)


CBIR, descriptor combination, similarity functions.


CBIR (Content Based Image Retrieval) systems are usually based on a series of distance functions, whose results are combined to produce similarity values which attempt to emulate the user’s judgements. Most of these distance functions produce dissimilarity values according to a given vector of features. Although the values they produce are related to the perceptual similarity between pictures, they usually do not have an easily describable meaning. In this paper, we propose a technique to normalize distance func-tions so that they express probability values. This normalization makes it possible to define a consistent method to combine several distance functions and produce a composite similarity measure which performs better than the original functions. To test the approach, a composite similarity measure has been built from a set of three statistically independent distance functions. Measuring the precision on the first 8 retrieved images for a 100 queries, the composite measure has outperformed the individual distance functions by 29%, and the normalized sum by about 14%.

Important Links:

Go Back