Abstract: A technique for the enhancement of degraded document images to improve their display quality characteristics and image recognition accuracy. Images believed to be representative of the same symbol which occur in different positions over a image source (e.g., a facsimile page) are clustered together. Using the symbols within a particular cluster, an average character image outline for that cluster of symbols is derived and thereafter used to refine the matching of symbols within the cluster and to determine a final representative symbol for that cluster. The final representative symbols from the various resulting clusters are then used to replace all matching images throughout the image source. Advantageously, the display quality and recognition accuracy of the image source is enhanced after application of the present invention due to the resulting improvement of the images in the image source.

Abstract: In a decision-forest classifier in accordance with the invention, a decision forest including multiple decision trees is used to classify "seen" training data and "unseen" data. Each individual tree performs an initial classification based on randomly selected subsets of the data. The classification outcomes by the individual trees are combined using a discriminant process in the decision-forest classier to render the ultimate classification decision.

Abstract: An image classifier, which performs character recognition, receives input images and assigns each input image to one of a plurality of image classes. The image classifier includes plural class distribution maps, each based on a plurality of features evaluated on training images, and each representing those feature values that occur at least once among the training images belonging to the corresponding class. The image classifier further includes means for constructing a test map by evaluating the plurality of features on the input image. The image classifier further includes means for comparing the test map to the class distribution maps in order to identify which one of the class distribution maps has the least distance to the test map. At least one of the features is defined according to a rule that relates to the shapes of images of at least one image class.