Document zone content classification and its performance evaluation

This paper describes an algorithm for the determination of zone content type of a given zone within a document image.We take a statistical based approach and represent each zone with 25 dimensional feature vectors. An optimized decision treeclassifier is used to classify each zone into one of nine zone content classes. A performance evaluation protocol is proposed.The training and testing datasets include a total of 24, 177 zones from the University of Washington English Document Imagedatabase III. The algorithm accuracy is 98.45% with a mean false alarm rate of 0.50%. SummaryA document can be divided into zones on the basis of its content. For example, a zone can be either text or non-text. Giventhe segmented document zones, correctly determining the zone content type is very important for the subsequent processeswithin any document image understanding system. This paper describes an algorithm for the determination of zone typeof a given zone within an input document image. In our zone content classification algorithm, zones are represented as 25dimensional feature vectors. The feature vector includes our new signature-like background analysis structure which is goodto study statistical characteristic of a given zone. A decision tree classifier is used to classify each zone into one of nineclasses on the basis of its feature vector. The protocol by which the decision tree classifier is optimized eliminates the dataover-fitting problem. To enrich our probabilistic model, we incorporate context constraints for certain zones within theirneighboring zones. We model zone class context constraints as a Hidden Markov Model and use Viterbi algorithm to obtainoptimal classification results. The training, pruning and testing datasets for the algorithm include a total of 1600 imagesfrom the University of Washington English Document Image database III. With a total of 24, 177 zones within the data set, across-validation method was used in the performance evaluation of the classifier. The classifier is able to classify each givenscientific and technical document zone into one of nine classes, 2 text classes (of font size 4− 18pt and font size 19− 32 pt),math, table, halftone, map/drawing, ruling, logo, and others. Using our zone content classification performance evaluationprotocol, the algorithm accuracy is 98.45% with a mean false alarm rate of 0.50%.