Robert J. Whitrow

Word shape analysis for a hybrid recognition system

This paper describes two wholistic recognizers developed for use in a hybrid recognition system. The recognizers use information about the word shape. This information is strongly related to word zoning. One of the recognizers is explicitly limited by the accuracy of the zoning information extraction. The other recognizer is designed so as to avoid this limitation. The recognizers use very simple sets of features and fuzzy set based pattern matching techniques. This not only aims to increase their robustness, but also causes problems with disambiguation of the results. A verification mechanism, using letter alternatives as compound features, is introduced. Letter alternatives are obtained from a segmentation based recognizer coexisting in the hybrid system. Despite some remaining disambiguation problems, wholistic recognizers are found capable of outperforming the segmentation based recognizer. When working together in a hybrid system, the results are significantly higher than that of the individual recognizers. Recognition results are reported and compared.

Multiple word segmentation with interactive look-up for cursive script recognition

Cursive script recognition is commonly based on finding letters within a word and recognizing them separately. The segmentation process is ambiguous and difficult. A method which combines word segmentation and letter recognition with lexical look-up in order to cope with segmentation ambiguity is presented. Words are first segmented into small elements which are then put together using a database of their possible combinations to produce alternative segmentations. Letter recognition is performed on each letter candidate and lexical look-up is applied, interactively, to prune illegal word recognition results. Lexical look-up is used to postulate word endings for partially recognized words. This provides the means of catering for words with sloppy endings, some misspellings and recovering from some recognition errors. An online cursive script recognition system, based on the above method, is described and evaluated.<<ETX>>

Recognising letters in on-line handwriting using hierarchical fuzzy inference

The recognition of unconstrained handwriting has to cope with the ambiguity and variability of cursive script. Preprocessing techniques are often applied to on-line data before representing the script as basic primitives, resulting in the propagation of errors introduced during pre-processing. This paper therefore combines pre-processing of the data (i.e. tangential smoothing) and encoding into primitives (Partial Strokes) in a single step. Finding the correct character at the correct place (i.e. letter spotting) is the main problem in non-holistic recognition approaches. Many cursive letters are composed of common shapes of varying complexity that can in turn consist of other subshapes. In this paper, we present a production rule system using Hierarchical Fuzzy Inference in order to exploit this hierarchical property of cursive script. Shapes of increasing complexity are found on a page of handwriting until letters are finally spotted. Zoning is then applied to verify their vertical position. The performance of letter spotting is compared with an alternative method.

Intelligent hierarchical layout segmentation of document images on the basis of colour content

This paper proposes a general methodology for automatic layout segmentation of documents. We first use colour histograms for extracting dominant colours of an image. This information is then used to hierarchically segment documents into regions of interest represented as polygons. If a region of interest is a picture the algorithm intelligently refrains from segmenting it further, while coloured regions that contain text are subsegmented. The method has been tested on 50 real life documents, such as office letters, brochures, and technical papers, scanned at 100/spl times/100 dpi resolution. Regions are detected with about 68% reliability. A critical analysis of the results is presented.

Recognizer characterisation for combining handwriting recognition results at word level

The paper concentrates on the combination of results of multiple recognizers at the word level. Two approaches are presented: word list merging and linear combination. Word list merging requires no knowledge about the individual recognizers. The linear combination is an attempt to exploit the information about characteristics of individual recognizers. This appears more complex than in the case of combination of results at the character level. Recognition of words is influenced by more factors, which can independently affect the recognition process. Characterisation of recognizers, used for word level combination, is more complex and requires more than a simple consideration of recognition success and failure. The concept of handwriting data characterisation is defined. A number of handwriting characteristics are extracted and used to guide the combination process. The choice of characteristics is made in the context of recognition methods used. No attempt at general characterisation of handwriting is made. The relationship between handwriting characteristics and recognition results is observed and used to obtain characteristics of individual recognizers. Results of the two combination methods are reported and compared with another frequently used method for results combination, the Borda count.

Recognition of facsimile documents using a database of robust features

A method for the recognition of poor quality documents containing touching characters is presented. The method is based on extraction of independent and robust features of each object of a sample word, where objects consist of single letters or of several touching ones. Thus avoiding letter segmentation the method eliminates errors frequently introduced in segmentation based approaches. Features are attributed by their position and extent in order to facilitate discrimination between different classes of objects. A method for automatic construction of a comprehensive database is presented. From a given dictionary every possible letter combination is obtained and the images of the artificially touching letters created. These images are subjected to noise and their features extracted. For recognition, alternatives for each object are found based on the database. Object alternatives are then combined into valid word alternatives using lexicon lookup. It has been observed that the developed method is effective for the recognition of poor quality documents.

DART-a software architecture for the creation of a Distributed Asynchronous Recognition Toolbox

Automatic analysis and recognition of documents is a complex task. Substantial research effort has therefore been spent addressing sub-tasks of manageable size. The solutions, however, have to be integrated into a comprehensive engine. Following the idea of a toolbox a software architecture has been developed that enables various tools to be combined into a network of cooperating recognition algorithms. Distributing the tools on a number of processors improves overall performance by parallel execution. The configuration of the tools as well as the connections between them can easily be modified even at run-time. Simple, interactive access to the various parameters and intermediate results of an algorithm aids the development and evaluation of experimental systems. The modular approach also facilitates extension of existing systems by the integration of novel methods. DART, the Distributed Asynchronous Recognition Toolbox architecture, realises these features. DART has been successfully used for recognition of handwriting and poor quality facsimiles.

Zoning invariant holistic recognizer for hybrid recognition of handwriting

The paper describes a holistic recognizer developed for use in a hybrid recognition system. The recognizer uses information about the word shape. As this information is strongly related to word zoning, care is taken to avoid limitations resulting from the inaccuracy of zone detection. The recognizer uses a very simple set of features and a fuzzy set based pattern matching technique. This aims to increase its robustness, but also causes problems with disambiguation of the results. A verification mechanism, using letter alternatives as compound features, is introduced. The letter alternatives are obtained from a segmentation based recognizer coexisting in the hybrid system. The holistic recognizer is found capable of outperforming the segmentation based one, despite the remaining disambiguation problems. When working together in a hybrid system, the results are significantly higher than those of the individual recognizers. Recognition results are reported and compared.

A Combined High and Low Level Approach to Interpreting Scanned Engineering Drawings

The effective computer interpretation of engineering drawing remains a desirable aim yet it continues to provide academic challenge. Much early work was concerned with the interpretation of low level vectorised data. For simple drawings, direct association and interpretation of the low level data often provides a very effective technique but drawing data, whether linework or higher level textual information, can be subject to inaccuracies and uncertainties of interpretation. Thus drawing errors and problems introduced by scanning are likely to introduce ambiguities which cannot be resolved directly from the low level data. The approach described in this paper combines features of a low level approach based on node and vertex association with a higher level interpretation of the textual content of the drawing. The textual description of dimensions, etc. has previously been used by the authors and by others, for the correction of drawing structures, in some cases using 3-D reconstruction as a means of validating the data association. The present work attempts to model an aspect of human drawing interpretation, whereby an ‘envelope of expectation’ is developed, through the interpretation of dimensioning and annotation information. This approach allows a link to be established between the highest level information on the drawing (such as the title block) and the low level vectors of the three elevations. It is thus no longer necessary to interpret obscure detail within the vector data directly. Separation of text on the drawing using OCR techniques allows the field of interpretation for the linework to be significantly narrowed.

Word recognition from sparse graphs of letter candidates using wildcards and multiple experts

Abstract Variability in handwriting styles suggests that many letter recognition engines cannot correctly identify some hand-written letters of poor quality at reasonable computational cost. Methods that are capable of searching the resulting sparse graph of letter candidates are therefore required. The method presented here employs ‘wildcards’ to represent missing letter candidates. Multiple experts are used to represent different aspects of handwriting. Each expert evaluates closeness of match and indicates its confidence. Explanation experts detemvne the degree to which the word alternative under consideration explains extraneous letter candidates. Schemata for normalisation and combination of scores are investigated and their performance compared. Hill-climbing yields near-optimal combination weights that outperform comparable methods on identical dynamic handwriting data.