Sang-Cheol Park

Text locating from natural scene images using image intensities

In this paper, we propose three text extraction methods based on intensity information for natural scene images. The first method is composed of gray value stretching and binarization by an average intensity of the image. This method is appropriate to extract texts from complex backgrounds. The second method is a split and merge approach which is one of well-known algorithms for image segmentation. The third one is a combination of the two. Experimental results show that the proposed approaches are superior to conventional methods both in simple and complex images.

Text localization in natural scene images by mean-shift clustering and parallel edge feature

A new text localization method using the parallel edge feature of text strokes is proposed, based on the observation that text-stroke consists of two edges in parallel. First, mean-shift clustering is employed to group similar pixels into clusters. The connected components in each cluster are considered as candidates for text strokes. Then, parallel edges are detected to verify whether the connected components are text strokes. The contribution of this paper is the presentation of a new feature of parallel edges along the stroke, providing structural information for the text localization. The performance, evaluated on ICDAR2003 image database, shows that the proposed algorithm works successfully with most of the text images.

Detection and tracking of multiple moving objects in video sequence using entropy mask method and fast level set method

In this paper, we propose a novel algorithm for real-time detection and tracking of multiple moving objects, which sequentially integrates the entropy difference method with adaptive threshold and the fast level set method. First, we have applied the Clausius Entropy theory to convert the pixel value in image domain into the amount of energy change in entropy domain. And then we apply the entropy difference detection method to detect the coarse region of the moving objects in this image and we have constructed the mask covering a detected coarse region. Second, taking the initial value of the level set for moving object as the constructed mask region, we have applied the fast level set technique to track rapidly the contour of detected objects. Here, we have used the fast level set method that combines the Fast Marching Method and the Smart Narrow Band. Experiment results demonstrate that our method can detect and track effectively and accurately the motion objects in video sequence.

An algorithm for candidate sequencing in non-dystrophic skeletal muscle channelopathies

Human skeletal muscle channelopathies (HSMCs) are a group of heritable conditions with ion channel–related etiology and similar presentation. To create a comprehensive picture of the phenotypic spectrum for each condition and to devise a strategy that facilitates the differential diagnosis, we collected the genotype and phenotype information from more than 500 previously published HSMC studies. Using these records, we were able to identify clear correlations between particular clinical features and the underlying alteration(s) in the genes SCN4A, CACNA1S, KCNJ2, and CLCN1. This allowed us to develop a clinical, symptom-based, binary decision flow algorithm that predicts the proper genetic origin with high accuracy (0.88–0.93). The algorithm was implemented in a stand-alone online tool (“CGPS”—http://cgps.ddd.co.kr) to assist with HSCM diagnosis in the clinical practice. The CGPS provides simple, symptom-oriented navigation that guides the user to the most likely molecular basis of the presentation, which permits highly targeted genetic screens and, upon confirmation, tailored pharmacotherapy based on the molecular origin.

Word extraction from table regions in document images

This paper describes a method to extract words from table regions in document images. The proposed approach consists of two stages: cell detection and word extraction. In the cell detection module, a table frame is extracted first by analyzing connected components and then intersection points are detected by a method using masks in the table frame. We correct false intersections, and detect the location of the cells within the table. In the word extraction module, a text region in each cell is located by using the connected components information that was obtained during the cell extraction module, and segmented into text lines by using projection profiles. Finally we divide the segmented lines into words using gap clustering and special symbol detection. The method correctly included character components touching the table frame with words, so experimental results show that more than 99% of words were successfully extracted from table regions.

A Keyword Matching for the Retrieval of Low-Quality Hangul Document Images

It is a difficult problem to use keyword retrieval for low-quality Korean document images because these include adjacent characters that are connected. In addition, images that are created from various fonts are likely to be distorted during acquisition. In this paper, we propose and test a keyword retrieval system, using a support vector machine (SVM) for the retrieval of low-quality Korean document images. We propose a keyword retrieval method using an SVM to discriminate the similarity between two word images. We demonstrated that the proposed keyword retrieval method is more effective than the accumulated Optical Character Recognition (OCR)-based searching method. Moreover, using the SVM is better than Bayesian decision or artificial neural network for determining the similarity of two images.

Bi-directional probabilistic hypergraph matching method using Bayes theorem

Establishing correspondences between two hyper-graphs is a fundamental issue in computer vision, pattern recognition, and machine learning. A hyper-graph is modeled by feature set where the complex relations are represented by hyperedges. Hence, a match between two vertex sets determines a hyper-graph matching problem. We propose a new bidirectional probabilistic hyper-graph matching method using Bayesian inference principle. First, we formulate the corresponding hyper-graph matching problem as the maximization of a matching score function over all permutations of the vertexes. Second, we induce an algebraic relation between the hyper-edge weight matrixes and derive the desired vertex to vertex probabilistic matching algorithm using Bayes theorem. Third, we apply the well known convex relaxation procedure with probabilistic soft matching matrix to get a complete hard matching result. Finally, we have conducted the comparative experiments on synthetic data and real images. Experimental results show that the proposed method clearly outperforms existing algorithms especially in the presence of noise and outliers.

Segmentation for medical image using a statistical initial process and a level set method

In this paper, we present a segmentation method for medical image based on the statistical clustering technique and the level set method. The segmentation method consists of a pre-processing stage for initialization and the final segmentation stage. First, in the initial segmentation stage, we adopt the Gaussian mixture model (GMM) and the Deterministic Annealing Expectation Maximization (DAEM) algorithm to compute the posterior probabilities for each pixels belonging to some region. And then we usually segment an image to assign each pixel to the object with maximum posterior probability. Next, we use the level set method to achieve the final segmentation. By using the level set method with a new defined speed function, the segmentation accuracy can be improved while making the boundaries of each object much smoother. This function combines the alignment term, which makes a level set as close as possible to a boundary of object, the minimal variance term, which best separates the interior and exterior in the contour and the mean curvature term, which makes a segmented boundary become less sensitive to noise. And we also use the Fast Matching Method for re-initialization that can reduce the computing time largely. The experimental results show that our proposed method can segment exactly the synthetic and CT images.

Text image spotting using local crowdedness and hausdorff distance

This paper investigates a Hausdorff distance, which is used for measurement of image similarity, to see whether it is also effective for document image retrieval. We proposed a method using a local crowdedness algorithm and a modified Hausdorff distance which has an ability of detection of partial text image in a document image. We found that the proposed method achieved a reliable performance of text spotting on postal envelops.

Keyword Spotting on Hangul Document Images Using Character Feature Models

In this Paper, we propose a keyword spotting system as an alternative to searching system for poor quality Korean document images and compare the Proposed system with an OCR-based document retrieval system. The system is composed of character segmentation, feature extraction for the query keyword, and word-to-word matching. In the character segmentation step, we propose an effective method to remove the connectivity between adjacent characters and a character segmentation method by making the variance of character widths minimum. In the query creation step, feature vector for the query is constructed by a combination of a character model by typeface. In the matching step, word-to-word matching is applied base on a character-to-character matching. We demonstrated that the proposed keyword spotting system is more efficient than the OCR-based one to search a keyword on the Korean document images, especially when the quality of documents is quite poor and point size is small.