Abstract :
We consider here the problem of image classification when more than one visual feature are available. In these cases, Bayes fusion offers an attractive solution by combining the results of different classifiers (one classifier per feature). This is a general form of the so-called "naive Bayes" approach. Analyzing the performance of Bayes fusion with respect to a Bayesian classifier over the joint feature distribution, however, is tricky. On the one hand, it is well-known that the latter has lower bias than the former, unless the features are conditionally independent, in which case the two coincide. On the other hand, as noted by Friedman, the low variance associated with naive Bayes estimation may dramatically mitigate the effect of its bias. In this paper, we attempt to assess the tradeoff between these two factors by means of experimental tests on two image data sets using color and texture features. Our results suggest that (1) the difference between the correct classification rates using Bayes fusion and using the joint feature distribution is a function of the conditional dependence of the features (measured in terms of mutual information), however: (2) for small training data size, Bayes fusion performs almost as well as the classifier on the joint distribution.
