Yves Rangoni

Document logical structure analysis based on perceptive cycles

This paper describes a Neural Network (NN) approach for logical document structure extraction. In this NN architecture, called Transparent Neural Network (TNN), the document structure is stretched along the layers, allowing an interpretation decomposition from physical (NN input) to logical (NN output) level. The intermediate layers represent successive interpretation steps. Each neuron is apparent and associated to a logical element. The recognition proceeds by repetitive perceptive cycles propagating the information through the layers. In case of low recognition rate, an enhancement is achieved by error backpropagation leading to correct or pick up a more adapted input feature subset. Several feature subsets are created using a modified filter method. The first experiments performed on scientific documents are encouraging.

Semi-automatic ground truth generation using unsupervised clustering and limited manual labeling

For training supervised classifiers to recognize different patterns, large data collections with accurate labels are necessary. In this paper, we propose a generic, semi-automatic labeling technique for large handwritten character collections. In order to speed up the creation of a large scale ground truth, the method combines unsupervised clustering and minimal expert knowledge. To exploit the potential discriminant complementarities across features, each character is projected into five different feature spaces. After clustering the images in each feature space, the human expert labels the cluster centers. Each data point inherits the label of its clusters center. A majority (or unanimity) vote decides the label of each character image. The amount of human involvement (labeling) is strictly controlled by the number of clusters - produced by the chosen clustering approach. To test the efficiency of the proposed approach, we have compared, and evaluated three state-of-the art clustering methods (k-means, self-organizing maps, and growing neural gas) on the MNIST digit data set, and a Lampung Indonesian character data set, respectively. Considering a k-nn classifier, we show that labeling manually only 1.3% (MNIST), and 3.2% (Lampung) of the training data, provides the same range of performance than a completely labeled data set would.

Labelling logical structures of document images using a dynamic perceptive neural network

This paper proposes a new method for labelling the logical structures of document images. The system starts with digitised images of paper documents, performs a physical layout analysis, runs an OCR and finally exploits the OCR’s outputs to find the meaning of each block of text (i.e. assigns labels like “Title”, “Author”, etc.). The method is an extension of our previous work where a classifier, the perceptive neural network, has been developed to be an analogy of the human perception. We introduce in this connectionist model a temporal dimension by the use of a time-delay neural network with local representation. During the recognition stage, the system performs several recognition cycles and corrections, while keeping track and reusing the previous outputs. This dynamic classifier allows then a better handling of noise and segmentation errors. The experiments have been carried out on two datasets: the public MARG containing more than 1,500 front pages of scientific papers with four zones of interest and another one composed of documents from the Siggraph 2003 conference, where 21 logical structures have been identified. The error rate on MARG is less than 2.5% and 7.3% on the Siggraph dataset.

Structure Extraction in Printed Documents Using Neural Approaches

This paper addresses the problem of layout and logical structure extraction from image documents. Two classes of approaches are first studied and discussed in general terms: data-driven and model-driven. In the latter, some specific approaches like rule-based or formal grammar are usually studied on very stereotyped documents providing honest results, while in the former artificial neural networks are often considered for small patterns with good results. Our understanding of these techniques let us to believe that a hybrid model is a more appropriate solution for structure extraction. Based on this standpoint, we proposed a Perceptive Neural Network based approach using a static topology that possesses the characteristics of a dynamic neural network. Thanks to its transparency, it allows a better representation of the model elements and the relationships between the logical and the physical components. Furthermore, it possesses perceptive cycles providing some capacities in data refinement and correction. Tested on several kinds of documents, the results are better than those of a static Multilayer Perceptron.

Data categorization for a context return applied to logical document structure recognition

The purpose of this work is to develop a pattern recognition system simulating the human vision. A transparent neural network, with context returns is used. The context returns consist in using global vision to correct local vision (i.e. input data are corrected according to neural network outputs). In order not to compute all the input features during these context returns, a filter-based method was designed to organize the features in clusters. This allows finding a good subset of input features during each cycle, which reduce the computations. The method interest is shown in the case of logical document structure retrieval.

XML Data Representation in Document Image Analysis

This paper presents the XML-based formats ALTO, TEI, METS used for digital libraries and their interest for data representation in a document image analysis and recognition (DIAR) process. In the first part we briefly present these formats with focus on their adequacy for structural representation and modeling of DIAR data. The second part shows how these formats can be used in a reverse engineering process. Their implementation as a data representation framework will be shown.