This paper presents a system for the offline recognition of large vocabulary unconstrained handwritten texts. The only assumption made about the data is that it is written in English. This allows the application of Statistical Language Models in order to improve the performance of our system. Several experiments have been performed using both single and multiple writer data. Lexica of variable size (from 10,000 to 50,000 words) have been used. The use of language models is shown to improve the accuracy of the system (when the lexicon contains 50,000 words, error rate is reduced by ∼50 % for single writer data and by ∼25 % for multiple writer data). Our approach is described in detail and compared with other methods presented in the literature to deal with the same problem. An experimental setup to correctly deal with unconstrained text recognition is proposed. Models.

A system for off-line cursive script recognition is presented. A new normalization technique (based on statistical methods) to compensate for the variability of writing style is described. The key problem of segmentation is avoided by applying a sliding window on the handwritten words. A feature vector is extracted from each frame isolated by the window. The feature vectors are used as observations in letter-oriented continuous density HMMs that perform the recognition. Feature extraction and modeling techniques are illustrated. In order to allow the comparison of the results, the system has been trained and tested using the same data and experimental conditions as in other published works. The performance of the system is evaluated in terms of character and word (with and without lexicon) recognition rate. Results comparable to those of more complex systems have been achieved.

This paper presents a surveyon o#-line Cursive WordRecogM]OyEL The approaches to the problem are described in detail. Each step of the processleading from raw data to the #nal result is analyzed. This survey is divided into two parts, the #rst onedealing with thegey,Hz aspects of Cursive WordRecog[zMyEL the second onefocusing on the applications presented in the literature. ? 2002 PatternRecogySzSk Society. Published by Elsevier Science Ltd. AllrigOL reserved. Ke5ti9tz Survey; O#-line cursive wordrecogHO[yEL Handwriting recogiting 1.

Contents 1 Introduction 5 2 State of the art 7 2.2 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.3 Structure of a CWR System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3.1 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.3.2 The segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.3.3 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.3.4 Lexicon reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.3.5 The data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3.6 The recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.3.7 Human Reading Inspired Systems . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.3.8 Holistic approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A system is developed for extracting stroke information from a grayscale, handwritten word. A temporal ordering of stroke segments is constructed, and this information is used to repair the skeleton produced by the thinning algorithm used. While currently not at optimum performance, much opportunity exists for further research in this field. Contents 1 Introduction 2 1.1 Background : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 2 1.2 Recognition Process : : : : : : : : : : : : : : : : : : : : : : : : : 3 1.3 Previous Work : : : : : : : : : : : : : : : : : : : : : : : : : : : : 4 1.4 Proposed System : : : : : : : : : : : : : : : : : : : : : : : : : : : 5 1.5 Implementation Notes : : : : : : : : : : : : : : : : : : : : : : : : 5 1.6 Organisation of Report : : : : : : : : : : : : : : : : : : : : : : : : 5 2 Thresholding and Thinning 6 2.1 Definitions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6 2.2 Thresholding and Filtering : : : : : : : : : : : : : : : ...

Handwriting Recognition generally uses a dictionary : in the case of large vocabularies, our system helps checking the existence of the words provided by the recognizer, or when there is a model to go through per word, it will reduce the number of models to go through. In the case of smaller vocabularies it can be used as a recognizer. From the signal provided by a digitizing tablet we build a global silhouette representing a word in a very simple and very compact way. After having made a statistical study about the possible silhouettes of each character in the alphabet and their frequency, we use these to build all the possible silhouettes to associate with each lexicon word and their corresponding probabilities. The lexicon is organized as an "epigenetic" neural network. In the recognizing phase, when retrieving words corresponding to a given silhouette, this network will let us reach not only the words having been describred with this silhouette, but also those possessing...

this paper, two methods are developed for determining the global slope of the word image. The first method estimates the slope by finding the minimum entropy in the vertical projection histograms, whereas the second method determines the slope from the frequency domain. 1) Minimum Entropy of Distribution: In this case, the dominant slope of the word is found from the slope corrected word which gives the minimum entropy of a vertical projection histogram. The grey level image is first binarised using the method developed by Otsu [37]. The vertical projection histogram (Figure 9) is calculated by counting the number of foreground pixels in each column of the binary image. The distribution is then normalised to have an area = 1. The basic idea can be demonstrated using a vertical line as an example. When the line is slanted at an angle, it will have a low flat distribution. When the line is upright, the distribution will be tall and narrow, which will result in a lower entropy measure than for the low flat distribution of the slanted line. The vertical projection histogram is calculated for a range of slope correction angles, ff i , where the angle is given in \SigmaR. A slope correction range of R = 60

Introduction Today, handwriting recognition is one of the most challenging tasks and exciting areas of research in computer science. Indeed, despite the growing interest in this field, no satisfactory solution is available. The difficulties encountered are numerous and include the huge variability of handwriting such as inter-writer and intra-writer variabilities, writing environment (pen, sheet, support, etc.), the overlap between characters, and the ambiguity that makes many characters unidentifiable without referring to context. Owing to these difficulties, many researchers have integrated the lexicon as a constraint to build lexicon-driven strategies to decrease the problem complexity. For small lexicons, as in bank-check processing, most approaches are global and consider a word as an indivisible entity [1] - [5]. If the lexicon is large, as in postal applications (city name or street name recognition) [6] - [10], one cannot consider a word as one entity, because of the huge num

This paper describes an elegant method for evaluating the discriminant power of features in the framework of an HMM-based word recognition system. This method employs statistical indicators, entropy and perplexity, to quantify the capability of each feature to discriminate between classes without resorting to the result of the recognition phase. The HMMs and the Viterbi algorithm are used as powerful tools to automatically deduce the probabilities required to compute the abovementioned quantities.