A novel segmentation based approach is proposed for recognition of offline handwritten Devanagari words. Stroke based features are used as feature vectors. A hidden Markov model is used for recognition at pseudocharacter level. The word level recognition is done on the basis of a string edit distance.

The fundamental representation of handwriting is a quasi-contiguous set of strokes. In offline settings handwriting stroke extraction is complicated by adjacent and overlapping strokes. The paper describes a reconstruction of the stroke information using a newly developed XYTC transform. This representation is very appropriate for handwriting feature extraction, stroke segmentation, underline removal, cross-out detection, and word recognition. Applications related to financial document processing are described.

This paper presents application of fusion techniques in a handwritten numeral recognition. The authors present two methods of fusing classifiers based on the fuzzy integral for handwritten numeral recognition. We also demonstrate that modification of fuzzy density calculation will result in better recognition rate. Test conducted on a large set of handwritten numeral data yielded high recognition rate with the proposed technique.

An unconstrained Farsi handwritten word recognition system based on fuzzy vector quantization (FVQ) and a hidden Markov model (HMM) for reading city names in postal addresses is presented. Preprocessing techniques including binarization, noise removal, slope correction and baseline estimation are described. Each word image is represented by its contour information. The histogram of chain code slopes of the image strips (frames), scanned from right to left by a sliding window, is used as feature vectors. Fuzzy c-means (FCM) clustering is used for generating a fuzzy code book. A separate HMM is trained by a modified Baum-Welch algorithm for each city name. A test image is recognized by finding the best match (likelihood) between the image and all of the HMM work models using a forward algorithm. Experimental results show the advantages of using an FVQ/HMM recognizer engine instead of conventional discrete HMMs.

Handwritten words are usually slant or italicized due to the mechanism of handwriting and the personality. In order to improve the accuracy of character segmentation and recognition, Kimura et al. (1993) proposed a chain code method for the slant estimation and correction. However the method is very simple and usually gives good estimate of the word slant, there was a problem such that the slant tends to be underestimated when the absolute of the slant is close or greater than 45/spl deg/. To solve the problem, we proposed an iterative chain code method. We introduce a noniterative method using 8-directional chain code for improving the linearity and the accuracy of the slant estimation. The experimental results show that the proposed method improves the linearity and the accuracy of the slant estimation efficiently without sacrificing the processing speed and the simplicity.

A holistic system for the recognition of handwritten Farsi/Arabic words using right-left discrete hidden Markov models (HMM) and Kohonen self-organizing vector quantization is presented. The histogram of chain-code directions of the image strips, scanned from right to left by a sliding window, is used as feature vectors. The neighborhood information preserved in the self-organizing feature map (SOFM), was used for smoothing the observation probability distributions of trained HMMs. Experiments carried out on test samples show promising performance results.

Handwritten words are usually slanted or italicized due to the mechanism of handwriting and one's personality. The authors have proposed a chain code method for slant estimation and correction to improve the accuracy of character segmentation and recognition. However, the method is very simple and usually gives a good estimate of the word slant but the relationship between the actual slant and the estimated slant is nonlinear, and the slant tends to be underestimated when the absolute of the slant is close or greater than 45/spl deg/. To resolve this, the paper proposes new methods for evaluating and improving the linearity and accuracy of slant estimation. The result of experimental study shows that the proposed algorithms are improved in the linearity and the accuracy of the slant estimation, and that a high speed iterative method is also efficient as far as processing time is concerned.

Recognition of license plate images is the topic of the paper. The major issue in this problem is the accurate extraction of the license plate string from varying backgrounds. The images are also characterized by non-uniform illumination. Recognition of the string is fairly straightforward, if the extraction process has been correctly designed. The authors discuss three different approaches to extraction and their study revealed that fusion of gray scale morphology and homomorphic processing yielded accurate extraction of the string. Recognition studies with 700 images captured during day and night times indicated an overall acceptance rate of 80% or higher with less than 1% confusion is achievable. Results obtained from a more exhaustive study that is currently in progress are presented.

In this paper, a technique for extracting the distortion parameters in filled-in forms is presented. The technique determines the transformations that is required to convert a filled-in form to match a known master and then extracts filled-in information. The method involves determining corresponding lines and keypoints between the master and the filled-in form and using the correspondence to determine the appropriate transformation. The correspondence problem is solved using results from geometric hashing voting scheme.

Handwritten character recognition by human readers, a statistical classifier, and a neural network is compared to know the required accuracy for handwritten word recognition. Sample characters extracted from postal address words on mail pieces collected by USPS were used to evaluate human and machine performance. Experimental results show that: 1) when the characters are segmented from words and are randomly presented, the accuracy of the machine recognition is comparable with the average human recognition accuracy, 2) the neural network employing the feature vector of size 64 outperforms the statistical classifier employing the same feature vector, and that 3) the statistical classifier employing the feature vector of size 400 achieves comparable recognition rate with the best human reader.

In this paper, a technique for extracting the distortion parameters in filled-in forms is presented. The technique determines the transformations that are required to convert a filled-in form to match a known master and then extracts filled-in information. The method involves determining corresponding lines and key points between the master and the filled-in form and using the correspondence to determine the appropriate transformation. The correspondence problem is solved using results from affine geometry.

The paper discusses the relative merits and complexities of two word recognition algorithms: lexicon directed and lexicon free techniques. This algorithm operates on a pre-segmented word image and yields the optimum concatenation of the image segments for each word in the lexicon. However, the computational complexity of this algorithm is quite high, as the optimum concatenation is required for every word in the lexicon. In the lexicon free word matching process, the character likelihood for all the letters are calculated and the maximum likelihood value and the associated letter are determined. In this approach an optimum string results from the concatenation process. The word matching process is applied only once for an input word image. Comparative results with regard to accuracy, speed and size of lexicon are presented.

A new algorithm for the extraction of words from printed documents that have interference marks and other strokes, cutting across text, is presented. Morphological operations based on multiple direction projection planes and skeleton images are adopted here to prevent the "flooding water" effect of conventional morphological operations. Test results indicate the feasibility of the proposed approach.

A technique for extracting filled-in information in form documents is presented. The transformation that is required to convert a filled-in form to match the master (blank form) is derived using results from projective geometry. Experimental studies with several forms indicate the proposed technique to be accurate and robust.

In this paper we propose a method to extract user entered components from various forms that may contain not only ruled lines but also gray-level background. The proposed technique utilizes a pair of form images: i) a blank form image and ii) a filled-in form image. Basically the pair of form images differ only at specific regions of the form reserved for insertion of handwritten (or machine printed) information. A sophisticated technique of inter image subtraction can extract the user entered components. In general, the background patterns of the reference image and sample images are not identical same because of printing and scanning conditions. The authors propose a morphological inter image subtraction to extract handwritten components from the filled-in forms. The user entered components are extracted by this method in a very robust manner.

In this paper, an evidence fusion technique, based on the notion of fuzzy integral is utilized to combine (fuse) the results of multiple character classifiers and realize a robust algorithm for high accuracy handwritten character recognition. Both source (classifier) relevance as well as source evidence are utilized to achieve significant enhancements. An algorithm for dynamically assigning source relevance, using the performance matrices of individual classifiers has also been developed. Experiments on a large data set show that a very low error rate and a low rejection rate can be achieved by fusing several simple classifiers.

The paper describes a new approach to ZIP code recognition using a word recognition algorithm, where a numeral string is recognized as a word. The paper also describes an end to end ZIP code recognition system consisting of tilt/slant correction, line segmentation, word segmentation, ZIP code location, as well as the ZIP code recognition. Evaluation tests are performed using address block image samples collected from United States mail pieces. The results of isolated numeral recognition, manually extracted ZIP code recognition, and end to end ZIP code recognition are presented to show and discuss the advantage of the word recognition based numeral string recognition.

This paper deals with the extraction of words from printed documents that have interference and other strokes (enclosing curves, underlines etc.). A new algorithm which combines the two processes of thinning and detection of directional strokes to effectively segment not only straight lines but also smooth arbitrarily oriented curves. The algorithms preserve both connectivity and directional characteristics. The morphological operations based on multiple direction projection planes and skeleton images are adopted here to successfully prevent the "flooding water" effect of conventional morphological operations. The effectiveness of the proposed algorithms is demonstrated on several experimental images.

In this paper, a technique for registering filled-in forms is presented. The technique determines the transformations that is required to convert a filled-in form to match a known master and then extracts filled-in information. This method involves determining corresponding points between the master and the filled-in form and using this correspondence to determine the appropriate transformation. The correspondence problem is solved using results from projective geometry.

The authors have researched two-letter state name (state name abbreviation) recognition and full state name recognition. According to this research, they think that the accuracy of the character segmentation is essential to recognize the word correctly, and it depends on the normalization of the word image. The normalization includes smoothing, underline removal, spurious blob removal, slant and base line correction etc. They present a new base line correction algorithm for the off-line handwritten words, which include cursive (continuous or running) words and hand-printed words. It uses background region analysis with the lower convex hull which is background area closed for three directions (upper, right, left), and the upper and bottom profiles of the merged convex hull. The authors show that the new method of base line correction is very powerful for most word images for city names of the USPS mail address database. The resulting image is useful for the holistic approach, and it's effective even when the image includes parts under the base line, for example, "f", "g", "j", or a very large character.

A digitized image that consists of text strings and uniformly distributed background symbols must be segmented if the characters in the text string are to be recognized. This paper describes the development and implementation of morphological approach to character string extraction from regular periodic overlapping text/background images that minimizes the shape distortion of characters. The effectiveness of this algorithm is demonstrated on several test images.>

Multiple expert system is shown to be a promising strategy for handwritten numeral recognition. This paper presents a multiple expert system using neural networks. In the proposed system, the authors have developed (1) an incremental clustering neural network algorithm with merging and canceling process, (2) a modified directional histogram feature extraction method and (3) a subclass method with learning rejection neuron strategy. Our experimental results on a large set of data show the efficiency and robustness of the proposed system.>

Discusses improvements made to a lexicon directed algorithm for recognition of unconstrained handwritten words (cursive, discrete, or mixed) such as those encountered in mail pieces. The procedure consists of binarization, pre-segmentation, intermediate feature extraction, segmentation recognition, and post-processing. The segmentation recognition and the post-processing are repeated for all lexicon words while the binarization to the intermediate feature extraction are applied once for an input word. The result of performance evaluation using large handwritten address block database is described, and algorithm improvements are described and discussed, in order to achieve higher recognition accuracy and speed. As a result the performance for lexicons of size 10, 100, and 1000 are improved to 98.01%, 95.46%, and 91.49% respectively. The processing speed for each lexicon is improved to 2.0, 2.5, and 3.5 sec/word on a SUN SPARC station 2.>

A new discrimination function for segmenting touching characters based on both pixel projection and profile projection is presented. A dynamic recursive segmentation algorithm is developed for effectively segmenting touching characters. Contextual information and a spelling checker are used to correct errors caused by incorrect recognition and segmentation. The 99.85% top recognition accuracy has been achieved, while the minimum accuracy is 99.4% based on 12 real documents.>

Proposes a new splitting algorithm for two-letter state name (state-name abbreviation) segmentation to recognize the state name on mail pieces. The merit of this algorithm is that the splitting point is located near the center of 0-projections, and it is especially effective for images in which the widths of each latter are different, such as "IN", "HI", "WI" and "MI". The performance of the algorithm is tested by real USPS mail envelopes. The correct segmentation rate obtained by the segmentation algorithm is 95.4% (91.7% by Kimura's splitting method) for two-letter state name images (932 images).>

Queueing formulations are presented for trading nodal storage capacity (SC) with the incremental transmission capacity (TC) of links in supra-high-speed (i.e., >10/sup 8//bps) packet-switched communication networks. The performance criterion used is the mean nodal forwarding time. The results are useful for designing all-optical packet-switching telecommunication networks where the designer needs to compute the amount of incremental TC (ITC) needed to eliminate a specified number of nodal buffers (SC) while preserving the mean nodal forwarding time.>

Detection of calcification in mammograms is analyzed based on mathematical morphology. A method of adaptively thresholding gray-level mammogram images is discussed. The threshold set is controlled by the index numbers in the skeleton of shapes which represent calcifications in mammograms. The parameters of the adaptive threshold sets are obtained from interpreting umbra shadows from an image function. The objective is to extract the suspicious areas from a mammogram and provide location information on certain predefined shapes and sizes of calcifications to radiologists for further examination.>

Presents the results obtained in a performance evaluation study on the recognition of handwritten numerals, using a statistical classifier and a neural network. The features consist of direction sensitive spatial features derived from the digitized image of the numeral. Studies with a large numeral database indicate that the overall performance of the neural net is better than that of the statistical classifier.>

The application of moment invariants to shape and pattern recognition problems is discussed. Zernike moment invariants and pseudo-Zernike moment invariants are derived in a new generalized form for the nth order. A new normalization procedure which improves the discrimination power of both Zernike and pseudo-Zernike moment invariants for handwritten numerals is also presented.>

A novel algorithm for the determination of surface shapes from intensity measurements is presented. The proposed model allows the estimation of the normals at any point on the surface and of the depth at any point in the image plane along the viewer (camera) direction. Simple shape descriptors for characterizing most developable surfaces are derived. It is shown that the principal curvatures of the surface are simply eigenvalues of the restriction of the shape matrix to the tangent space. The proposed algorithm allows incorporation of prior information obtained from stereopsis and other vision processes. Tests with intensity images have shown the algorithm to be convergent, robust, and accurate.>

An interpolation of the sparse depth map of stereopsis is a necessary accomplishment in early vision. A neural implementation of a class of interpolants, known as the Shepard surfaces, is discussed. The computing nature of neurons is exploited, and the results suggest that a variety of early visual functions could be neurally realized in this fashion.>

Algorithms are presented that are capable of restoring images degraded by a variety of corrupting noise sources. These include any combination of additive and/or multiplicative Gaussian noise, focus blur and nonlinear distortion. A parallel implementation of the algorithm is shown to be feasible.>

Restoration of noisy images through the use of a Gibbs field model for the image and a Gaussian random field characterization of the corrupting noise is discussed in this paper. Two new algorithms capable of parallel implementation are presented and shown to yield satisfactory restoration of images corrupted by high levels of noise.

An algorithm for multilevel image segmentation is presented. The main feature is that it is based on repetition of a binary segmentation algorithm, in a fashion similar to the binary expansion of an integer number. The binary segmentation considered assumes a Markov Random Field model for the original scene, and an additive i.i.d, noise signal. Simulation rasults are presented, and compared with other filtering algorithms, such as median filtering.

This paper discusses the results of some experiments that were performed to test the feasibility of high speed vector quantization scheme for low bit rate speech coding. Three different methods of codebook generation are presented and their performance as evaluated by the average distortion, signal to quantization noise, speed of implementation are discussed.

A new set of topological features (primitives) for use with a syntactic classifier for high-accuracy recognition of handwritten numerals is proposed. The tree grammar used in this study makes it possible to achieve high-recognition speeds with minimal preprocessing of the test pattern.

This paper introduces a new concept of the VLSI-compatible multiprocessor system to perform most of digital signal processing functions, required in speech processing and modern telecommunications network. High performance is achieved by high degree of concurrency and parallelism in the architecture of the processor. Identified attributes of signal processing have influenced substantially the architecture. Set of macro instructions has been defined to be stored in the local memory inside the DSP chip. This approach reduces the access time and speeds-up overall operations. The key features of the DSP processor are presented. The multiprocessor architecture is described and possible configurations are discussed.

A survey of research efforts in the area of speaker recognition indicate that for the same choice of speaker-dependent speech parameters the recognition accuracy is significantly affected by the distance measure used. In this work several distance classifiers are evaluated for use in text-independent speaker identification. The four distance measures investigated are the Mahalanobis distance, maximum a posteriori probability, nearest neighbor criterion and the correlation distance measure. It is found that both the maximum a posteriori probability criterion and the correlation distance measure yield extremely poor results. The Mahalanobis distance and the nearest neighborhood criterion yield relatively poor results (error\sim20-30%) with the former consistently superior to the latter. It is shown that these scores can be improved through a proposed variation of the nearest neighbor method.

This work addresses the development of a reliable, high accuracy text-independent speaker recognition system for a small population, with the reference parameters characterizing each speaker obtained from short segments of speech. Initially the potential for speaker discrimination of several different vocal parameter sets was investigated. These included the LPC, Reflection, Cepstrum and Log Area Ratio coefficients, speech power spectrum parameters and the inverse filter spectral coefficients. It was then decided to use any two parameter sets in a composite decision-making scheme. A "repeat feature" was incorporated into the speaker recognition system, whereby a speaker was asked to read a fresh test speech segment if the decisions made by using the two different parameter sets individually were not coincident. Test results indicate that a significant improvement in accuracy is realizable.

This paper discusses the development of an efficient, general method for the design of 2-D recursive digital filters. It effectively uses the 1-D filter theory. The technique employs the McClellan Transformation (MT). Methods are suggested for the design of McClellan Transformation Coefficients (MTC). A stability error criteria is formed on the basis of Complex Cepstrum (CC) and it guarantees the stability of the filter. Group delay and magnitude characteristics are further refined using nonlinear optimization. The method can also be used for stabilizing unstable filters. Numerical implementation and results are presented with an example.

This paper deals with noise reduction in images using two-dimensional Kalman filtering. To reduce the computational load and the memory storage, updating of state vector is done within a certain distance of the point currently being processed, the support of the filter and dynamical model is restricted to a non-symmetric half plane. Direct histogram specification method is combined with Kalman filtering algorithm. Significant improvement in the quality of the image is obtained. The technique used is particularly attractive for on-line applications. It is also shown that the various gradient computations and template matching schemes work efficiently on estimated image to extract vital features (edges) from the image for the future analysis and studies. The results obtained confirm the feasibility of these algorithms.

The main objective of this work was to investigate the effectiveness of long-term averages of the orthogonal linear prediction parameters in text-independent speaker recognition. To investigate the possibility of feature selection, a technique using dynamic programming (1) was used to select a subset of k best features among the entire set N. The results indicate that the parameters comprising the optimal set chosen are speaker-dependedt. Verification accuracies of 96.5% were obtained using the selected optimal 8- parameter (out of 12) feature set for each speaker in a verification scheme, in which the reference parameters were generated from 100 seconds of time-spaced voiced speech and the test parameters were generated from 5 seconds of voiced speech.

The goal of this study was to develop a speaker verification algorithm which yielded high accuracy verification when test utterances were obtained under different noise conditions. In this study, test utterances were obtained i) from a noisy telephone channel ii) with additive white noise iii) from recordings in rooms with typical ambient noise. Initially the verification algorithm utilizing orthogonal parameters derived from time normalized speech without preprocessing were used to evaluate the accuracies obtainable. Our studies indicated that accuracies were limited to about 45%. The test utterances obtained above were then processed by the adaptive noise cancelling algorithm to filter out the distortion. The verification algorithm, when applied to the processed speech yielded a significant improvement in verification accuracy to the extent of about 95%.

An important problem in speaker verification systems arises, when the speech inputs are noise corrupted with signal to noise ratios in the range of 10 to 30 dB (noise assumed to be zero mean and white). This paper deals with the accuracy of speaker verification algorithms derived from an orthogonal parameter representation of speech. Initially, the investigations are directed to evaluate the sensitivity of orthogonal parameters to the level of noise in the speech signal. The accuracy of verification is then determined, using only those parameters that are least sensitive to additive noise. The influence of the order of the linear prediction model on verification is also studied. The verification algorithm is based on the distance measure used by Sambur. Finally thresholds are established to determine the proper choice of orthogonal parameters (used in distance computation) and the order of the linear prediction model for a given signal to noise ratio in the speech signal. The above study is then used to evaluate the accuracy of verification when speech inputs are obtained from a noisy telephone channel.