In surveillance applications, search space reduction (SSR) is an essential element to efficient algorithms. In this study, spatial and temporal SSRs are integrated for license plate detection in video sequences; the plates could be extracted robustly and extremely fast. Our method started from spatial SSR by a bi-level one-pass plate extraction (BOPE) algorithm developed to extract plates accurately even in complicated situations. Moreover, we proposed to exclude repeated patterns with the similar appearances in the same location of consecutive frames, which usually include stopped vehicles or regular backgrounds. For efficiency, repeated patterns were detected only on the results of BOPE, named spatiotemporal SSR, based on a block-based mechanism by estimating the tangent distance, which is invariant to the variations in positions, sizes, rotations, or brightness. To reduce the computational load, the repeated patterns and their measured invariant features could be retained for next estimation. In our experiments, the search space could be reduced up to 87.9% by the spatiotemporal SSR; all plate numbers with the heights ranging from 24 to 40 are extracted correctly. The high performance, average 76 frames per second on a 3G Hz PC, suggests that the real-time goal for surveillance applications could be accomplished.

This paper describes a plate recognition system that can process images rapidly at high accuracy rates. This system is designed to meet the requirements of performance, computational speed, and adaptation for vehicle surveillance applications, such as stolen car detection systems. These requirements are satisfied by adopting a cascade framework, utilizing plate characteristics, and developing fast one-pass algorithms. Our system is composed of three main cascading modules for plate detection, character segmentation, and postprocessing. Each module is further decomposed into several cascading procedures, which are composed of successively more complex rejecters. The first module rapidly rejects a majority of nonplate regions by using low computational gradient features and a one-pass scanning algorithm followed by heavy computational statistical rejecters. The second module rejects a majority of noncharacter regions in a similar manner. A peak-valley analysis algorithm is proposed to rapidly detect all promising candidates of character regions. The third module eliminates the plate characters that do not satisfy the plate specifications. In our experiments, the system can recognize plates over 38 frames per second with a resolution of 640 times 480 pixels on a 3-GHz Intel Pentium 4 personal computer

The existence of characters with various font families or italic face will decrease the performance of character recognition significantly. This paper aims to improve the performance of a multi-font Chinese character recognition system. We first obtain the angle in italic face characters. If the characters of a text line belong to the italic face, we straighten the characters in the text line. Then we determine the font family of the text line. If the family is Ming-font, which has more distinct features than other fonts to be extracted, we recognize the text line by the Ming-font kernel. In post-processing, we use font family information to adjust the recognition results and select the most promising recognition candidate. Experimental results show the effectiveness of the proposed system.

In this paper, we present an automatic understanding system for Chinese name cards. After preprocessing of an input card image, we grouped characters into item blocks and segmented characters according to their aspect ratios and gap widths. After character extraction, we sent characters to a statistical multi-font character recognizer. We identified items according to their geometric characteristics and embedded key-characters. A user can edit segmentation results interactively and then save the results in the database for further applications. The high performance in the experiments shows the effectiveness of the proposed system.

To discover translation knowledge in diverse data resources on the Web, this article proposes an effective approach to finding translation equivalents of query terms and constructing multilingual lexicons through the mining of Web anchor texts and link structures. Although Web anchor texts are wide-scoped hypertext resources, not every particular pair of languages contains sufficient anchor texts for effective extraction of translations for Web queries. For more generalized applications, the approach is designed based on a transitive translation model. The translation equivalents of a query term can be extracted via its translation in an intermediate language. To reduce interference from translation errors, the approach further integrates a competitive linking algorithm into the process of determining the most probable translation. A series of experiments has been conducted, including performance tests on term translation extraction, cross-language information retrieval, and translation suggestions for practical Web search services, respectively. The obtained experimental results have shown that the proposed approach is effective in extracting translations of unknown queries, is easy to combine with the probabilistic retrieval model to improve the cross-language retrieval performance, and is very useful when the considered language pairs lack a sufficient number of anchor texts. Based on the approach, an experimental system called LiveTrans has been developed for English--Chinese cross-language Web search.

This article presents an approach to automatically extracting translations of Web query terms through mining of Web anchor texts and link structures. One of the existing difficulties in cross-language information retrieval (CLIR) and Web search is the lack of appropriate translations of new terminology and proper names. The proposed approach successfully exploits the anchor-text resources and reduces the existing difficulties of query term translation. Many query terms that cannot be obtained in general-purpose translation dictionaries are, therefore, extracted.

This paper presents a novel binarization algorithm for color document images. Conventional thresholding methods do not produce satisfactory binarization results for documents with close or mixed foreground colors and background colors. Initially, statistical image features are extracted from the luminance distribution. Then, a decision-tree based binarization method is proposed, which selects various color features to binarize color document images. First, if the document image colors are concentrated within a limited range, saturation is employed. Second, if the image foreground colors are significant, luminance is adopted. Third, if the image background colors are concentrated within a limited range, luminance is also applied. Fourth, if the total number of pixels with low luminance (less than 60) is limited, saturation is applied; else both luminance and saturation are employed. Our experiments include 519 color images, most of which are uniform invoice and name-card document images. The proposed binarization method generates better results than other available methods in shape and connected-component measurements. Also, the binarization method obtains higher recognition accuracy in a commercial OCR system than other comparable methods.

In this paper, we propose a method for calligraphy document binarization. First, the properties of calligraphy documents are analyzed to classify calligraphy documents into two classes: tablet documents and writing documents. Because calligraphy documents have serious noises and large variations in gray scales of the characters, the foreground will be broken after thresholding. The anisotropic diffusion technique is used to smooth input calligraphy documents, which can preserve character boundaries well. Last, noises are removed from binarized images based on the information of stroke widths. In our experiments, 90 tablet documents and 51 writing documents are binarized and well binarized results are obtained.

In this paper, we propose a method to vectorize Chinese characters in calligraphy documents. Our system can prevent the zigzag phenomena when the characters are enlarged. The system contains two modules: contour segment extraction and description. In the former, high curvature points on contours are detected as corner points, which divide the contour into several segments. In the latter, a contour segment can be described either by a straight line or a cubic Bezier curve. According to relations between the contour segment and the Bezier curve, control points are adjusted to fit the contour segment better. When the curve fitness cost is small enough, the shape is described well. The processing time of our curve fitting is about five seconds per A4 image, which has 4488 contour segments. Experimental results demonstrate that our system is efficient and promising.

This paper proposes a general Chinese document recognition system with high recognition rate, including preprocessing, recognition kernel, and post-processing, especially for low quality images. In the preprocessing module, fast rotation transformation algorithm is proposed. Since characters are extracted for recognition engines, document images must be segmented into text blocks, text lines, and then character images. In the recognition module, two recognition engines are used to recognize the character images. The weights of these kernels and features are calculated from the relative stroke widths of character images. In the post-processing module, we calculate confidence values for different candidates and then select the most confident candidate as the OCR result. The experiments show the system we propose is very effective and efficient.

The paper presents an approach to automatically extracting translations of Web query terms through mining of Web anchor texts and link structures. One of the existing difficulties in cross-language information retrieval (CLIR) and Web search is the lack of the appropriate translations of new terminology and proper names. Such a difficult problem can be effectively alleviated by our proposed approach, and the resource of anchor texts in the Web is proven a valuable corpus for this kind of term translation.

This paper presents an approach to automatically extracting the bilingual translations of many Web query terms through mining the Web anchor texts. Some preliminary experiments are conducted on using 109,416 Web pages containing both Chinese and English anchor texts in their in-links to extract Chinese translations of 200 English queries selected from popular query terms in Taiwan. It is found that the effective translations of 75% of the popular query terms can be extracted, in which 87.2% cannot be obtained in common translation dictionaries.

This paper presents a novel approach to grouping Chinese handwritten field-data of form documents using a gravitation-based algorithm. We develop an algorithm to extract handwritten field data which may be written out of the fields. We first extract and remove form lines for input form images. Next, we detect connected-components from remaining data, and compute the gravitation for each connected-component by using the black pixel counts as their mass. Then, we move connected-components to their field center according to their gravitation, since filled-in data have the locality property, that is, data of the same field are usually written in a local area consecutively. After moving these connected-components for a certain times, we can assign most components to the fields where they should be. Thus, we can determine which connected-components should be extracted for a particular field. Experimental results show that this proposed method can group field-data effectively.

Characters sometimes overlap with nontextual lines in form documents and these interfered-character would generally be recognized with poor accuracy. In this paper we propose a two-step interfering-line removing method. Positions and orientations of interfering-lines are first detected by the Hough transform. Interfering-lines widths are then determined from projection histograms. An ambiguous area is defined to bound an interfering-line. Black runs in the ambiguous are classified into four types and grouped into run-groups. The directions of hidden character strokes in each run-groups are predicted. Black pixels located in these hidden strokes are regarded as character pixels and the other black pixels are considered as interfering pixels, which will be removed then. Most OCR engines are trained by noninterfered sample characters. In order to recognize interfered-characters, we adjust the feature values by assigning a stability value to each of subregions. In our collected 1820 interfered handwritten Chinese characters, the recognition accuracy was 24.02% for interfered-characters and 89.91% for characters after removing interfering-lines.

This paper proposes a confusion set analysis approach to increasing character recognition rate. In the training phase, top five candidates are outputted from an OCR system. Each candidate is assigned a weight in reversed-rank. If the reversed-rank sum of an output candidate is greater than a threshold, the input character is stored in the confusion set of the output candidate. In the recognition phase, the characters in the confusion set with respect to the output character are recognized by another OCR system to find the final recognition result.

In this paper, we propose a methodology for identifying typefaces of printed Chinese characters in documents. Three kinds of features, stroke width means, stroke width variations, and aspect ratio, are first used to classify character typefaces as: Black, Li, Kai-Round, or Ming-Song. Each of the last two groups contains two typefaces. Vertical/horizontal stroke width ratios are used to distinguish between the Ming and Song typefaces and accumulative pixel ratio to distinguish between the Kai and Round typefaces. Six different typeface feature distributions measured from 5401 printed Chinese characters are considered, and a trapezoid-shaped membership function is constructed for each distribution. Based on these membership functions, we determine what typeface each input character belongs to using a two-level decision tree. To increase the identification rate, the typeface of a certain character is adjusted according to the typeface identification results of the front and the next characters. In the character recognition system, we use two statistical features: crossing counts and contour directional counts. We achieved an 89.87% typeface identification rate in our experiments, and a 95.60% character recognition rate.

Business cards include many kinds of information, such as names, addresses and telephone numbers. In order to use the information effectively, it is necessary to extract the information from the cards automatically in order to build a database. The goal of this paper is to extract and recognize characters from color business cards. To separate the foreground from the background in a card, we assign all pixels to eight color types. Then we calculate a dynamic threshold using the color information to extract the foreground. Next, we extract the characters by four steps: (i) connected component extraction, (ii) local thresholding, (iii) mark, line and noise deletion, and (iv) character grouping. Finally, we recognize the characters by a statistical Chinese and English character recognition system. We test 30 business cards which have Chinese characters, English characters, numerals and punctuation marks. The extraction rate and accuracy for our system are 96.97% and 95.43% respectively. The recognition rate is 88.78% for Chinese characters and 97.58% for English characters, numerals and punctuation marks.

We present two techniques for speeding up character recognition. Our character recognition system, including the candidate cluster selection and detail matching modules, is implemented using two statistical features: crossing counts and contour direction counts. In the training stage, we divide characters into different clusters. To keep a very high recognition rate, the candidate cluster selection module selects the top 60 clusters with minimal distances from among 300 predefined clusters. To further speed up the recognition speed, we use a modified branch and bound algorithm in the detail matching module. In the automatic document reading system, characters and punctuation marks are first extracted from printed document images and sorted according to their positions and the document orientation. The system then recognizes all printed Chinese characters between pairs of punctuation marks. The results are then spoken aloud by a speech synthesis system.

A form processing system aims to extract meaningful data from a form document for office automation. To locate the data, we have to extract and understand the form structure. In this paper, a strip projection method is presented for extracting form structure. We first segment the input form image uniformly into vertical and horizontal strips. Since most lines in a form are vertical and horizontal lines, we project the image in each vertical strip horizontally and in each horizontal strip vertically. The peak positions in the projection profiles denote the possible existence of lines in the form image. Next we trace the lines started from the possible line positions in the source image. After all lines are extracted, redundant lines are removed by a line verification algorithm and broken lines are linked by a line merging algorithm. This proposed method can reduce much computation time than other methods such as Hough transformation and line detection and approximation algorithm. Experimental results demonstrate that the proposed method is very effective.

This paper presents a new method for clustering the words in a dictionary into word groups, which are applied in a Chinese character recognition system with a language model to describe the contextual information. The Chinese synonym dictionary Tong2yi4ci2 ci2lin2 providing the semantic features is used to train the weights of the semantic attributes of the character-based word classes. The weights of the semantic attributes are next updated according to the words of the behavior dictionary, which has a rather complete word set. Then, the updated word classes are clustered into m groups according to the semantic measurement by a greedy method. The words in the behavior dictionary can finally be assigned into the m groups. The parameter space for bigram contextual information of the character recognition system is m/sup 2/. From the experimental results, the recognition system with the proposed model has shown better performance than that of a character-based bigram language model.

We present a system to segment and recognize texts and mathematical expressions in a document. The system can be divided into six stages: page segmentation and labeling, character segmentation, feature extraction, character recognition, expression formation, and error correction and expression extraction. In expression formation, we build a symbol relation tree for each text line to represent the relationships among the symbols in the text line. Some heuristic rules based on the primitive tokens are used to correct the recognition errors in a text line. We extract all mathematical expressions according to some basic expression forms. Our database consists of 190 symbols in the current stage. The average recognition rate is about 96.16%.

This paper presents a Chinese handwritten character generating system. Second-order online models in our database are used to describe 228 radicals. Twelve relations and a position evaluation function are used by the second-order online models. Forty types of strokes are defined by line vectors. All characters, radicals and strokes are represented by 1-D strings. The first order relation on the online model is used to construct radicals and the second-order relation is used to check if the constructed radical has correct stroke relation. A Turing-like test is used to determine whether computer generated characters are similar to human written characters. Experimental results show that almost 75% of generated characters are taken as handwritten characters by the testers. The average space for storing each character is 22 bytes, which occupies much less storage than characters generated by other systems.

A system for understanding mathematical expressions is presented. After separating all symbols in an input mathematical expression, we utilize thirteen features to represent each symbol. In order to reduce the computational time, a coarse classification algorithm is used to reduce the number of candidates. Then for each input symbol, the character with the highest similarity is selected as the candidate symbol. Since some of the symbols in an arithmetical expression may touch each other, a dynamic programming algorithm is adopted to identify correct characters from connected symbols. In the expression formation stage, a procedure-oriented method is proposed for translating the recognized symbols appearing in a 2D space into a 1D character string. The authors have used 105 mathematical expressions as training data and 50 expressions as testing data. The experimental results have demonstrated the feasibility of the understanding system.>

A two-stage Chinese text recognition system is presented. In the first stage, a Chinese character is first segmented nonuniformly into 10 strips horizontally and vertically. Then three statistical features, viz. crossing counts peripheral background area and contour line length are extracted to form a 60-dimension feature vector. A feature matching method based on the city-block distance metric is employed to select N nearest neighbors as the candidates for each input character from the reference template base, which consists of 5,401 frequently-used Chinese characters. In the second stage, a 3-part-of-speech (tri-POS) Markov language model is employed to extract the most promising characters from all candidate characters in an input sentence. The dynamic programming method is applied to find the most promising sentence hypothesis whose part-of-speech sequence has the maximum likelihood to occur among all of the candidate sentences for an input sentence. The tri-POS contextual information is estimated from a tagged corpus.>

A handwritten Chinese character generator that generates handwritten Chinese characters with different variations is proposed and used for evaluating the performance of an OCR system. The character generator first generates radicals of a character stroke by stroke and then combines the generated radicals to form a line-vector character. Then, the line-vector character is thickened to obtain a character image. The generated characters are then used to evaluate the stability of stroke extractors and calculate the peak performance of a matching module and the recognition error caused by stroke extractors.>

The authors propose a two-stage text recognition system with high recognition rate. In the first stage, characters recognized as different recognition results by two matching modules are rejected, since they are recognized incorrectly by at least one of the two matching modules. The rejected characters are then processed by a Markov language model in the second stage. Since most of the input characters are recognized in the first stage, the computation cost of the language model is low and the recognition rate of the language model is excellent. By using this erroneously-identified text recognition approach, a system with high character recognition accuracy can be achieved. Our text recognition system can be extended to a multiple-stage recognition system to further improve the recognition accuracy. In each stage, various matching modules can be used. The recognized result of an input character will be accepted only when all matching modules produce the same result. Rejected characters will be fed into the next stage for further processing.>

This paper proposes a probabilistic approach to recognize handwritten Chinese characters. According to the stroke writing sequence, strokes and interleaved stroke relations are built manually as a 1-D string, called an on-line model, to describe a Chinese character. In an input character, strokes are first extracted by a tree searching method. The recognition problem is then formulated as an optimization matching problem in a multistage directed graph, where the number of stages is the length of the modelled stroke sequence. Nodes in a stage represent extracted strokes that have the same stroke type as defined in the on-line model and the link between two neighboring nodes corresponds to the relationship between the two extracted strokes. The probability that the extracted stroke belongs to the predefined stroke type is calculated from the stroke line segments, and the transition probability between two extracted strokes is the degree of satisfaction of the relationship defined in the on-line model. The Viterbi algorithm, which can handle stroke insertion, deletion, splitting, and merging, is applied to recover the sequence of strokes consisting of the unknown character. The similarity is defined to be the product of stroke probabilities and stroke transition probabilities in the stroke sequence. The unknown character is matched with all modelled characters and is recognized as the one with the highest similarity. Experiments with 540 characters uniformly selected from the database CCL/HCCR1 (250 variations/class) are conducted, and the recognition rate is about 92.8%, which proves the feasibility of the proposed recognition system.

This paper presents a probabilistic approach to recognize handwritten Chinese characters. According to the stroke writing sequence, strokes and interleaved stroke relations are built manually as a 1D string, called online models, to describe a Chinese character. The recognition problem is formulated as an optimization process in a multistage directed graph, where the number of stages is the length of the modelled stroke sequence. Nodes in a stage represent extracted strokes. The Viterbi algorithm, which can handle stroke insertion, deletion, splitting, and merging, is applied to compute the similarity between each modelled character and the unknown character. The unknown character is recognized as the one with the highest similarity. Experiments with 500 characters uniformly selected from the database CCL/HCCR1 are conducted, and the recognition rate is about 94.3%.>

A computer vision method is presented to determine the 3-D spatial locations of joints or feature points of human body from a film recording the human motion during walking. The proposed method first applies the geometric projection theory to obtain a set of feasible postures from a single image, then it makes use of the given dimensions of the human stick figure, physiological and motion-specific knowledge to constrain the feasible postures in both the single-frame analysis and the multi-frame analysis. Finally a unique gait interpretation is selected by an optimization algorithm. Computer simulations are used to illustrate the ideas presented.>

An approach for determining the location and orientation of a camera mounted on an ALV (autonomous land vehicle) is proposed. The authors choose the road boundaries and the objects with vertical edges to be the calibration objects. During the calibration process, the ALV is supposed to move with pure translation. The camera models are derived by using two or three images under the assumptions that the height of the camera and the distance during taking images are known in advance. The tilt and swing angles can be computed from the first image. The pan angle can then be derived by using the projections of vertical edges in three images. If the angle between the moving direction of the ALV and the direction of road boundaries is large enough, two images are enough to find the pan angle. Two translational parameters can be found by employing the projections of vertical edges in two or three images. >