Abstract: This paper presents analyses of different methods of post-processing lines that have resulted from the raster-to-vector conversion of black and white line drawing. Special attention was paid to the borders of connected components of maps. These methods are implemented with compression and smoothing algorithms. Smoothing algorithms can enhance accuracy, so using both smoothing and compression algorithms in succession gives a more accurate result than using only a compression algorithm. The paper also shows that a map in vector format may require more memory than a map in raster format. The Appendix contains a detailed description of the new smoothing method (continuous local weighted averaging) suggested by the authors.

Abstract: Semantic analysis of cartographic images is interpreted as a separate representation of cartographic patterns (alphanumeric, punctual, linear, and area). We present an approach to map interpretation exploring the idea of synthesis of invariant graphic images at low level processing (vectorization and segmentation). This means that we ran "vectorization-recognition" and "segmentation-interpretation" systems simultaneously. Although these systems can generate some errors in interpretation, they are much more useful for the following understanding algorithms because its output is nearly recognized objects of interest.

Abstract: This paper describes a trainable recognizer for hand-drawn sketches using geometric features. We compare three different learning algorithms and select the best approach in terms of cost-performance ratio. The algorithms employ classic machine-learning techniques using a clustering approach. Experimental results show competing performance (95.1%) with the non-trainable recognizer (95.8%) previously developed, with obvious gains in flexibility and expandability. In addition, we study both their classification and learning performance with increasing number of examples per class.

Abstract: The Arc Segmentation Contest, as the fourth in the series of graphics recognition contests organized by IAPR TC10, was held in association with the GREC'2001 workshop. In this paper we present the extended summary of the contest: the contest rules, performance metrics, test images and their ground truths, and the outcomes.

Abstract: In this paper, we present a method for correcting a skeleton-based vectorization. The method robustlys egments the skeleton of an image into basic features, and uses these features to reconstruct analytically all the junctions. It corrects some of the topological errors usually brought byp olygonal approximation method, and improves the precision of the junction points detection.We first give some reminders on vectorization and explain what a good vectorization is supposed to be. We also explain the advantages and drawbacks of using skeletons. We then explain in detail our correction method, and show results on cases known to be problematic.

Abstract: This paper describes a method for the recognition of arcs and circles in engineering drawings and other scanned images containing linework. The approach is based on vectorizing a binary image, smoothing the vectors to a sequence of small straight lines, and then attempting to fit arcs. The software was successful in winning first prize in the arc segmentation contest held at GREC 2001, and the results are presented in the context of this evaluation.

Abstract: For small screen devices, such as PDAs, which totally depend on a pen-based user interface, traditional menu-selection/button-clicking based user interface becomes inconvenient for graphics inputting. In this paper, a novel sketch-based graphics inputting user interface is presented. By sketching a few constituent primitive shapes of the user-intended graphic object, the candidate graphic objects in the shape database are guessed and displayed in a ranked list according to their partial structural similarity to what the user has drawn. The user can then choose the right one from the list and replace the sketchy strokes with the exact graphic object with proper parameters, such as position, size and angle. This user interface is natural for graphics input and is especially suitable for schematic design.

Abstract: We examine the nature of ground-truth: whether it is always well-defined fora given task, oron ly relative and approximate. In the conventional scenario, reference data is produced by recording the interpretation of each test document using a chosen data-entry platform. Looking a little more closely at this process, we study its constituents and their interrelations. We provide examples from the literature and from our own experiments where non-trivial problems with each of the components appear to preclude the possibility of real progress in evaluating automated graphics recognition systems, and propose possible solutions. More specifically, for documents with complex structure we recommend multi-valued, layered, weighted, functional ground-truth supported by model-guided reference data-entry systems and protocols. Mostly, however, we raise far more questions than we currently have answers for.

Abstract: This paper presents an ongoing project working on an optical handwritten music manuscript recognition system. A brief background of Optical Music Recognition (OMR) is presented, together with a discussion on some of the main obstacles in this domain. An earlier OMR prototype for printed music scores is described, with illustrations of the low-level pre-processing and segmentation routines, followed by a discussion on its limitations for handwritten manuscripts processing, which led to the development of a stroke-based segmentation approach using mathematical morphology. The pre-processing sub-systems consist of a list of automated processes, including thresholding, deskewing, basic layout analysis and general normalization parameters such as the stave line thickness and spacing. High-level domain knowledge enhancements, output format and future directions are outlined.

Abstract: The user interface is critical to the success of a diagram recognition system. It is difficult to define precise goals for a user interface, and even more difficult to quantify performance of a user interface. In this paper, we discuss some of the many research questions related to user interfaces in diagram recognition systems. We relate experiences we have gathered during the construction of two on-line diagram recognition systems, one for UML (Unified Modeling Language) notation and the other for mathematical notation. The goal of this paper is to encourage discussion. The graphics recognition community needs strategies and criteria for designing, implementing, and evaluating user interfaces.

Abstract: In this paper, we propose our new method to recognize matrices including repeat symbols and area symbols The method consists of 4 parts; detection of matrices, segmentation of elements, construction of networks and analysis of the matrix structure In the construction of networks, we regard a matrix as a network of elements connected each other by links representing their relative relations, and consider its horizontally projected network and vertically projected one In the analysis, we obtain the areas of variable block pattern elements generating the minimum rectangular area of the matrix by solving the simultaneous system of equations given by the two projected networks We also propose a format to represent the structure of matrices to output the result of the matrix recognition

Abstract: We present in this paper how to apply to mathematical formulae a generic recognition method already used for musical scores, table structure and old forms recognition We propose to use this method to recognize the structure of formulae and also to recognize some symbols made of line segments This offers two possibilities: improving the symbol recognition when there is a lot of symbols like in mathematics; and overcoming segmentation problems we usually find in old mathematical formulae

Abstract: Programs for understanding hand-drawn sketches and diagrams must interpret curvilinear configurations that are sloppily drawn and highly variable in form. We propose a two-stage subgraph matching framework for sketch recognition that can accommodate great variability in form and yet provide efficient matching and easy extensibility to new configurations. First, a rectification stage corrects the initial data graph for the common deviations of each kind of constituent local configuration from its ideal form. The model graph is then matched directly to the data by a constraint-based subgraph matching process, without the need for complex error-tolerance. We explore the approach in the domain of human stick figures in diverse poses.

Abstract: This paper briefly describes an experimental arc extraction algorithm that ran the Arc Segmentation Contest at GREC'2001. As the proposed method is based on the one detailed in [5], this paper only describes the improvments we brought to the original method. We first review some rules from the evaluation protocol that helped us to make major assumptions while designing the algorithm. We then explain the method, and discuss the results we obtained in various cases. Finally, we give some conclusions and introduce a possible extension to this method.

Abstract: This paper motivates an approach to graphics recognition grounded in a framework for human and machine vision known as Perceptual Organization. We review some of the characteristics of this approach that distinguish it from traditional engineering of document recognition systems, and we suggest why and how the techniques and philosophy of Perceptual Organization might lead to advances in the very practical matters of interpreting diagrams, drawings, and sketches.

Abstract: This work presents an approach to map interpretation starting from unstructured vector data. For this task a map interpreter based on PROLOG and grammatical object descriptions has been considered. The major challenge in this approach is the definition of production rules and grammars which are general enough to handle different data sets and which are specific enough to discriminate the different object types.

Abstract: Image understanding systems rely heavily on a priori knowledge of their application domain, often exploiting techniques developed in the wider field of knowledge-based systems (KBSs). Despite attempts, typified by the KADS/CommonKADS projects, to develop structured knowledge engineering approaches to KBS development, those working in image understanding continue to employ unstructured 1st generation KBS methods. We analyse some existing image understanding systems, concerned with the interpretation of images of line drawings, from a knowledge engineering perspective. Attention focuses on the relationship between the structure of the systems considered and the KADS/CommonKADS models of expertise, sometimes called generic task models. Mappings are identified between each system and an appropriate task model, identifying common inference structures and use of knowledge. This is the first step in the acquisition of models of the expertise underpinning drawing interpretation. Such models would bring significant benefits to the design, maintenance and understanding of line drawing interpretation systems.

Abstract: The main objective of this paper is to experiment a new approach to develop a high level document analysis platform by composing existing components from a comprehensive library of state-of-the art algorithms Starting from the observation that document analysis is conducted as a layered pipeline taking syntax as an input, and producing semantics as an output on each layer, we introduce the concept of a Component Algebra as an approach to integrate different existing document analysis algorithms in a coherent and self-containing manner Based on xml for data representation and exchange on the one side, and on combined scripting and compiled libraries on the other side, our claim is that this approach can eventually lead to a universal representation for real world document analysis algorithmsThe test-case of this methodology consists in the realization of a fully automated method for generating a browsable, hyper-linked document from a simple scanned image Our example is based on cutaway diagrams Cutaway diagrams present the advantage of containing simple "browsing semantics", in the sense that they consist of a clearly identifiable legend containing index references, plus a drawing containing one or more occurrences of the same indices

Abstract: This paper proposes a novel approach for extracting element clusters, associated with individual folding steps of origami, from a page-image of origami drill books. Unlike the documents investigated in other researches, origami books are usually not designed under a specific layout structure. Our approach is based on a Voronoi expression schema. Since the only clue to group related elements is spatial proximity relations among the elements, we utilize Voronoi regions, which are generated by using the centroids of main elements as characteristic points of Voronoi diagrams, to estimate the scopes of element clusters. This method provides necessary information to a recognition system that converts a folding process, described by a series of illustrations, instructive sentences and special symbols, into a 3D animation. The experimental results with good accuracy show the effectiveness of the proposed method.

Abstract: This paper describes an experimental setup, conducted in collaboration with the ISA research group of the LORIA laboratory, Oce-PLT, and students from the Ecole des Mines de Nancy. The main objective is to experiment an approach to develop a high level document analysis platform by composing existing bricks from a comprehensive library of state-of-the art algorithms. The test-case of this methodology consists in the realization of a fully automated method of generating a browsable, hyper-linked document from a simple scanned image. We concentrated our work on cutaway diagrams. These documents present the advantage of containing simple browsing semantics, in the sense that they consist of a clearly identifiable legend containing index references, plus a drawing containing one or more occurrences of the same indices. The setup described in this paper starts from a raw binary image of a cutaway diagram, and delivers an XML description matching the references of the legend with the indices in the image, and a browser for interpreting the XML generated map. The complete document treatment pipeline is conceived within a combined scripting and compiled library environment.

Abstract: Automating the interpretation of a map in order to locate some geographical objects and their relations is a challenging task, which goes beyond the transformation of map images into a vectorized representation and the recognition of symbols. In this work, we present an approach to the automated interpretation of vectorized topographic maps. It is based on the generation of logic descriptions of maps and the application of symbolic Machine Learning tools to these descriptions. This paper focuses on the definition of computational methods for the generation of logic descriptions of map cells and briefly describes the use of these logic descriptions in an inductive learning task.

Abstract: In this paper, we consider the general problem of technical document interpretation, applied to the documents of the French Telephonic Operator, France Telecom. At GREC'99, we presented a new set of features, based on the Fourier-Mellin transform (FMT), allowing a good classification of multi-oriented and multiscaled pattern in comparison with classical approach. For this GREC'01, we propose the use of this set of features for the rotation and scale parameters estimation, through the use of the shift theorem of the Fourier transform. A comparison with a parameter estimation issued from Zernike moments is also given.

Abstract: Document analysis and graphics recognition algorithms are normally applied to the processing of images of 2D documents scanned when flattened against a planar surface. Technological advancements in recent years have led to a situation in which digital cameras with high resolution are widely available. Consequently, traditional graphics recognition tasks may be updated to accommodate document images captured through a camera in an uncontrolled environment. In this paper the problem of document image rectification is discussed. The rectification targets the correction of perspective and geometric distortions of document images taken from uncalibrated cameras, by synthesizing new views which are better suited for existing graphics recognition and document analysis techniques. The proposed approach targets cases in which the document is not necessarily flat, without relaying on specific modeling assumptions, and by utilizing one or more overlapping views of the document. Document image rectification results are provided for several cases.

Abstract: In this paper, vectorization is achieved by the recognition of architectural drawing images To obtain vector components, an architectural drawing recognizer (ADR) is developed The ADR recognizes line components such as main wall, dimension line and points that indicate the dimension of scanned architectural drawing images A merging process from separately recognized line components is accomplished in order to increase the accuracy of result vectors The recognition ratio and the vectorization of ADR is 983% over 9 real scanned apartment building images

Abstract: Engineering solutions are generally documented in assembly and part drawings and bill of materials. A great benefit, qualitatively and commercially, can be achieved if these paper based storages can be transformed into digital information archives. The process of this transformation is called reconstruction. The reconstruction process of paper based assembly drawings consists of four steps: digitization; vectorization/ interpretation; 3D reconstruction of the parts and the 3D reconstruction of the assembly.This paper evaluates existing commercial systems worldwide for interpretation of paper based mechanical engineering drawings. For a complete reconstruction process a 3D reconstruction is needed. This functionality is already supported by some CAD systems to a certain extent, but it still remains a major topic of research work. One CAD system which converts 2D CAD models into 3D CAD models is presented. Finally, after the reconstruction of the parts the whole assembly can be reconstructed. Until now, no system for the automatic reconstruction of assemblies is available. In our paper we present a general approach for automatic reconstruction of 3D assembly model data by interpretation of mechanical engineering 2D assembly drawings, their part drawings, and the bill of materials.

Abstract: A new frame extraction and a removal method for processing form documents is proposed. The method robustly extracts scanned preprintings such as frames and lines. It consists of a frame detection process and a frame removal process. In the frame detection process, the center coordinates are extracted using a Generalized Hough Transformation-based method. Then, using those coordinates, an inscribed rectangular image for each frame is produced. In the frame removal process, the detected frame image is removed along the outside of the rectangular edge. These processes are repeated to remove the target frames successfully by changing some pre-processings such as reducing and enhancing. The method was applied to some types of images. They are postal codes on mail and forms received by facsimiles. In both cases, there often can be seen low quailty pre-printings. For those low quality images, convetional approach such as model pattern maching was not well worked because of local distortion. Through experiments in frame detection and removal of the images, we demonstrated that all of the frames could be successfully removed.

Abstract: In this paper we introduce a method for knowledge extraction in early stages of design by the interpretation of low-level CAD data, i.e. the geometry of mechanical products. Therefore, only the 3D CAD model of the product, including its components, is needed. The information about physical principles and functions is extracted automatically. First a more detailed introduction to the design process in mechanical engineering is presented. This enables the reader to understand the information which is accumulated during product design and which must be recognized during the interpretation process, i.e. the redesign. Then the approach of low-level CAD data interpretation is presented in detail. This includes the recognition of physical principles and product functions and the automatic construction of effective geometries and function structures. At the end the approach is verified by presenting an example of a product geometry, which is interpreted stepwise.

Abstract: An efficient method of classifying form is proposed in this paper. Our method identifies a small number of matching areas by their distinctive images with respect to their layout structure and then form classification is performed by matching only these local regions. The process is summarized as follows. First, the form is partitioned into rectangular regions along the locations of lines of the forms. The disparity in each partitioned region of the comparing form images is measured. The penalty for each partitioned area is computed by using the preprinted text, filled-in data, and the size of a partitioned area. The disparity and penalty are considered to compute the score to select final matching areas. By using our approach, the redundant matching areas are not processed and a feature vector of good quality can be extracted.

Abstract: The separation of overlapping text and graphics is a challenging problem in document image analysis. This paper proposes a specific method of detecting and extracting characters that are touching graphics. It is based on the observation that the constituent strokes of characters are usually short segments in comparison with those of graphics. It combines line continuation with the feature line width to decompose and reconstruct segments underlying the region of intersection. Experimental results showed that the proposed method improved the percentage of correctly detected text as well as the accuracy of character recognition significantly.