Multi-Class Document Layout Classification using Random Chopping

This paper proposes a multi-class document layout classification/recognition system using a method called random chopping. A scanned document image undergoes text line extraction and is represented as a set of quadrilaterals for every pair of text lines. For compact representation, a dictionary of quadrilateral clusters is built beforehand, and a document image is then represented as a word occurrence histogram by looking up its quadrilaterals in the dictionary. The training process iteratively chops all training classes into two partitions and trains a linear classifier for this split. A binary coordinate space is built from all chops, and every document’s histogram descriptor is then projected to this space to form a binary signature. Layout similarity is reduced to distance computation between two signatures. Our experiments demonstrate that this multi-class classification system achieves very good performance not only on trained classes but also on instances from layout classes never seen in the training.