Hsi-Jian Lee

Determination of 3D human body postures from a single view

Abstract In this paper a method is proposed to recover and interpret the 3D body structures of a person from a single view, provided that (1) at least six feature points on the head and a set of body joints are available on the image plane, and (2) the geometry of head and lengths of body segments formed by joints are known. First of all, the feature points on the head in the head-centered coordinate system and their image projections are used to determine a transformation matrix. Then, the camera position and orientations are extracted from the matrix. Finally, the 3D coordinates of the head points expressed in the camera-centered coordinate system are obtained. Starting from the coordinates of the neck, which is a head feature point, the 3D coordinates of other joints one-by-one are determined under the assumption of the fixed lengths of the body segments. A binary interpretation tree is used to represent the 2n − 1 possible body structures, if a human body has n joints. To determine the final feasible body structures, physical and motion constraints are used to prune the interpretation tree. Formulas and rules required for the tree pruning are formulated. Experiments are used to illustrate the pruning powers of these constraints. In the two cases of input data chosen, a unique or nearly unique solution of the body structure is obtained.

Binarization of color document images via luminance and saturation color features

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.

Knowledge-guided visual perception of 3-D human gait from a single image sequence

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. >

Anchor text mining for translation of Web queries: A transitive translation approach

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.

Detection and recognition of license plate characters with different appearances

This paper proposes an approach to developing an automatic license plate recognition system. Car images are taken from various positions outdoors. Because of the variations of angles from the camera to the car, license plates have various locations and rotation angles in an image. In the license plate detection phase, the magnitude of the vertical gradients is used to detect candidate license plate regions. These candidate regions are then evaluated based on three geometrical features: the ratio of width and height, the size and the orientation. The last feature is defined by the major axis. In the character recognition phase, we must detect character features that are non-sensitive to the rotation variations. The various rotated character images of a specific character can be normalized to the same orientation based on the major axis of the character image. The crossing counts and peripheral background area of an input character image are selected as the features for rotation-free character recognition. Experimental results show that the license plates detection method can correctly extract all license plates from 102 car images taken outdoors and the rotation-free character recognition method can achieve an accuracy rate of 98.6%.

Translation of web queries using anchor text mining

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.

Recognition of handwritten Chinese characters via short line segments

Abstract A system to recognize handwritten Chinese characters is presented. In the first stage, short line segments are extracted, where a new efficient algorithm is proposed, based on accumulated chain codes, for line approximation. In the feature extraction stage, features of each character are computed, which are further defined by the features of its composing line segments. In the matching stage, dynamic programming is first used to calculate the similarity between a segment of the input character and a segment of the reference character, and then the similarity between the two characters is computed. In order to reduce the number of candidates, a coarse classification algorithm is proposed. The recognition rate for 150 characters is about 90%, excluding the error in coarse classification.

Recognition-based handwritten Chinese character segmentation using a probabilistic Viterbi algorithm

Abstract This paper presents a recognition-based character segmentation method for handwritten Chinese characters. Possible non-linear segmentation paths are initially located using a probabilistic Viterbi algorithm. Candidate segmentation paths are determined by verifying overlapping paths, between-character gaps, and adjacent-path distances. A segmentation graph is then constructed using candidate paths to represent nodes and two nodes with appropriate distances are connected by an arc. The cost in each arc is a function of character recognition distances, squareness of characters and internal gaps in characters. After the shortest path is detected from the segmentation graph, the nodes in the path represent optimal segmentation paths. In addition, 125 text-line images are collected from seven form documents. Cumulatively, these text-lines contain 1132 handwritten Chinese characters. The average segmentation rate in our experiments is 95.58%. Moreover, the probabilistic Viterbi algorithm is modified slightly to extract text-lines from document pages by obtaining non-linear paths while gaps between text-lines are not obvious. This algorithm can also be modified to segment characters from printed text-line images by adjusting parameters used to represent costs of arcs in the segmentation graph.

Extraction and recognition of license plates of motorcycles and vehicles on highways

A recognition system is proposed to extract and recognize license plates of motorcycles and vehicles on highways. In the first stage, a block-difference method is used to detect moving objects. According to the variance and the similarity of the M/spl times/N blocks defined on two diagonal lines, the blocks are categorized as three kinds: low-contrast, stationary and moving blocks. In the second stage, a screening method based on the projection of edge magnitudes is used to find two peaks in the projection histograms to bound license plates. The scanning lines with low counts can be removed. In the third stage, character images are segmented and recognized. In our experiments, we tested 180 pairs of images. The block-difference method has a 98% success rate and can remove 88% of pixels from an image on average. The screening method has a 94.4% success rate and the character recognition method has a 95.7% precision rate.

Design of a mathematical expression recognition system

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%.