AUTOMATIC IMAGE ANNOTATION USING ADAPTIVE COLOR CLASSIFICATION

We describe a system which automatically annotates images with a set o f prespecified keywords, based on supervised color classification of p ixels into N prespecified classes using simple pixelwise operations. T he conditional distribution of the chrominance components of pixels be longing to each class is modeled by a two-dimensional Gaussian functio n, where the mean vector and the covariance matrix for each class are estimated from appropriate training sets. Then, a succession of binary hypothesis tests with image-adaptive thresholds has been employed to decide whether each pixel in a given image belongs to one of the prede termined classes. To this effect, a universal decision threshold is fi rst selected for each class based on receiver operating characteristic s (ROC) curves quantifying the optimum ''true positive'' vs ''false po sitive'' performance on the training set. Then, a new method is introd uced for adapting these thresholds to the characteristics of individua l input images based on histogram cluster analysis. If a particular pi xel is found to belong to more than one class, a maximum a posteriori probability (MAP) rule is employed to resolve the ambiguity. The perfo rmance improvement obtained by the proposed adaptive hypothesis testin g approach over using universal decision thresholds is demonstrated by annotating a database of 31 images. (C) 1996 Academic Press, Inc.