Structure detection and segmentation of documents using 2D stochastic context-free grammars

Abstract

In this paper we define a bidimensional extension of Stochastic Context-Free Grammars for structure detection and segmentation of images of documents. Two sets of text classification features are used to perform an initial classification of each zone of the page. Then, the document segmentation is obtained as the most likely hypothesis according to a stochastic grammar. We used a dataset of historical marriage license books to validate this approach. We also tested several inference algorithms for Probabilistic Graphical Models and the results showed that the proposed grammatical model outperformed the other methods. Furthermore, grammars also provide the document structure along with its segmentation.

Deep neural networks for record counting in historical handwritten documents

“…We tested the proposed approach on a benchmark dataset proposed by [5] for addressing the segmentation of historical handwritten documents. As previously discussed, one significant problem to train CNNs is the limited number of labeled samples in the dataset and in other collections available for research.…”

“…We therefore generated a synthetic dataset adopting the approach described in Section 2 obtaining a synthetic training set with 81,060 pages with associated information on the number of records in each page. In particular, the training set contains pages with a number of records comprised between 3 and 9 even if the benchmark collection in [5] contains only pages with 5,6, or 7 records each. When generating the training set we used the real images only to infer the structure of the pages and the overall structure of the records as well as to extract the page background.…”

“…In the first one, we used stratified cross validation to estimate the error rate on a larger dataset. In the second experiment we compared the results achievable by the proposed approach with those described in [5] considering the same splitting of the data in training, validation, and test datasets.…”

“…In the second experiment, we compared the results obtained by our system with those presented in [5]. In order to perform a fair comparison we used the splitting of training and test data as proposed in [5].…”

“…The record detection problem has been addressed in [5], where an EM-based layout analysis method is proposed. The approach is tested on a collection of marriage license books where each page contains a variable number of handwritten records.…”