This report is divided into two parts. In the first part we present an overview of sensor fusion. We analyze the single steps in a fusion process and describe in a systematic way several methods for combining pattern classifiers. The second part describes some practical experiments with classifier combination in the field of face recognition. We investigate the impact of decision combination on the recognition rate. Several combination strategies and their results are compared. The face classifiers used for the practical experiments are shortly presented, too. We used two full face classifiers (HMMs, eigenfaces) and a profile classifier (shape comparison). Thus, the combination is based on two completely different information sources. CR Categories and Subject Descriptors: I.4.8 [Image Processing]: Scene Analysis; I.5.0 [Pattern Recognition]: General; I.5.2 [Pattern Recognition]: Design Methodology. General Terms: Algorithms, Design. Additional Key Words: Sensor Fusion, Clas...

This paper presents a complex approach for the content-based text categorization of printed German business letters into pre-defined message types such as order, invoice, offer, etc. The categorization results of two competing classifiers are combined by means of a voting component embodying knowledge about the strengths and weaknesses of the classifiers. The individual classifiers differ strongly in their basic assumptions: While the first one considers layout and typographic information with respect to certain keywords the second one is a more conventional text categorization approach which merely incorporates textual features. Since this whole categorization tool is embedded into a document analysis system, a highly precise classification is essential for a subsequent goal-directed extraction of structured information aimed at the integration of the document into the current business workflow of a company

table of contents, OCR, journal splitting, text mining, text chunking, document understanding This paper introduces a novel journal splitting algorithm. It takes full advantage of various kinds of information such as text match, layout and page numbers. The core procedure is a highly efficient text-mining algorithm, which detects the matched phrases between the content pages and the title pages of individual articles. Experiments show that this algorithm is robust and able to split a wide range of journals, magazines and books.

Systems for postcorrection of OCR-results can be fine tuned and adapted to new recognition tasks in many respects. One issue is the selection and adaption of a suitable background dictionary. Another issue is the choice of a correction model, which includes, among other decisions, the selection of an appropriate distance measure for strings and the choice of a scoring function for ranking distinct correction alternatives. When combining the results obtained from distinct OCR engines, further parameters have to be fixed. Due to all these degrees of freedom, adaption and fine tuning of systems for lexical postcorrection is a difficult process. Here we describe a visual and interactive tool that semi-automates the generation of ground truth data, partially automates adjustment of parameters, yields active support for error analysis and thus helps to find correction strategies that lead to high accuracy with realistic effort.

This article presents a case study on the combination of classifiers for the recognition of handwritten digits. Four different classifiers are briefly described and evaluated using the NIST-digits data set. Different parallel and sequential combination schemes are introduced. Furthermore, it is described how to tune the sequential combination using a boosting technique. These combination methods are benchmarked using the NIST-digits. The experimental results indicate that all investigated classifier combinations outperform the best individual classifier. The sequential combination yields slightly better results than the parallel combination and has a much lower computational complexity. In addition, it is possible to improve the performance of the sequential combination by boosting. 1 Introduction Automatic handwriting recognition has a variety of applications at the interface between man and machine. In this article we will focus on off-line systems that are a key component for some ...

This paper describes the derivation of probability of correctness from scores assigned by most recognizers. Motivation for this research is three-fold: (i) probability values can be used to rerank the output of any recognizer by using a new set of training data; if the training data is sufficiently large and representative of the test data, the recognition rates are seen to improve significantly, (ii) derivation of probability values puts the output of different recognizers on the same scale; this makes comparison across recognizers trivial, and (iii) word recognition can be readily extended to phrase and sentence recognition because the integration of language models becomes straightforward. We have conducted an extensive set of experiments. The results show a reranking of recognition choices based on the derived probability values leading to an enhancement in performance. The performance of many different digit recognizers improved by 1-4% points on a blind set of images. 1 Introduct...

We introduce a multiple classifier system that incorporates a global optimization technique based on a Genetic Algorithm for dynamically selecting the set of experts to use in the majority vote approach. The proposed technique is applicable when the experts in the pool provide both the class assigned to the input sample and a measure of the reliability of the this classification. For each sample, the experts selected for participating in the majority vote are those whose reliability is larger than a given threshold. There are as many thresholds as the number of experts by the number of classes. The values of the thresholds aimed at selecting the best set of experts for each input sample are determined by a canonical Genetic Algorithm. The reliability measures provided by the experts of the pool are also used to implement the tie-break mechanism needed within the majority vote scheme. The system has been tested on a handwritten digit recognition problem, and its performance compared with those exhibited by other multiexpert systems exploiting different combining rules.

In this paper we propose a reject rule applicable to a Multi-Expert System (MES). The rule is adaptive to the given domain and allows the achievement of the best trade-off between reject and error rates as a function of the costs attributed to errors and rejects in the considered application. The results of the method are particularly effective since the method does not rely on particular statistical assumptions, as other reject rules. An experimental analysis carried out on publicly available databases is reported together with a comparison with other methods present in the literature.

Combination functions typically used in biometric identification systems consider as input parameters only the matching scores related to a single person in order to derive a combined score for that person. We present a systematic framework to use all scores received by all persons as input to a single combination function when sufficient training data is available. More fundamentally, we identify four types of classifier combination methods determined by number of combining functions that must be trained and the number of input parameters. We investigate in detail combination methods, which consider all available matching scores as input parameters to a single trainable combination function. We describe how such methods account for dependencies between scores output by any single participating classifier. Our experiments demonstrate the advantage of using such combination methods when dealing with large number of classes, as in the case with biometric person identification systems.

Traditional approaches to document object location use a single locator that is expected to locate as many instances of the object class of interest as possible. However, if the class includes subclasses with diverse visual characteristics or is not characterized by easily computable visual features, it is difficult for the single locator to account for wide variation in object characteristics within the class. As a result, increasingly complex models of objects to be located are used. An alternative approach is to combine the decisions of multiple locators, each of which is suitable for certain image conditions. This approach utilizes a collection of simple locators that complement one another, rather than relying on one complex locator. An effective method for combining the location results is vital to the success of this approach. This thesis presents a theory of combining the results of multiple document object locators tuned to different object characteristics. The purpose of the ...