Transductive learning of logical structures from document images

A fundamental task of document image understanding is to recognize semantically relevant components in the layout extracted from a document image. This task can be automatized by learning classifiers to label such components. The application of inductive learning algorithms assumes the availability of a large set of documents, whose layout components have been previously labeled through manual annotation. This contrasts with the more common situation in which we have only few labeled documents and an abundance of unlabeled ones. A further degree of complexity of the learning task is represented by the importance of spatial relationships between layout components, which cannot be adequately represented by feature vectors. To face these problems, we investigate the application of a relational classifier that works in the transductive setting. Transduction is justified by the possibility of exploiting the large amount of information conveyed in the unlabeled documents and by the contiguity of the concept of positive autocorrelation with the smoothness assumption which characterizes the transductive setting. The classifier takes advantage of discovered emerging patterns that permit us to qualitatively characterize classes. Computational solutions have been tested on document images of scientific literature and the experimental results show the advantages and drawbacks of the approach.