Selecting important genes from microarray data is a considerably challenging problem as shown in Guyon's 2002 paper in this journal. We have developed an alternative feature ranking and selection methodology to tackle this problem. On the basis of several cancer data sets, AMFES (adaptive multiple features selection) outperforms Guyon's RFE (recursive feature elimination). In this paper, we will present a comprehensive and systematic comparison of three methods: AMFES, RFE, and the CORR (correlation coefficient) on five data sets (leukemia, colon, lymphoma, prostate, and potentially others). The leukemia, colon, and lymphoma data sets are adapted from Guyon's paper for convenience and the prostate cancer data set is from a public database, NCBI GEO (Gene Expression Omnibus). These three methods are compared in terms of test accuracy, number of selected features, computational time (total and training), statistical significance (t test, p values, and ROC (receiver operating characteristic)/AUC (area under curve)), and the discovery rate of informative features. AMFES obtains better results in computational time and number of selected features while maintaining higher or comparable test accuracy, statistical significance, and the discovery rate of informative features. In addition, AMFES can serve as a general methodology for other similar problems such as sampling and data mining.

In this paper, we propose a method for classifying textual entities of bilingual documents written in Chinese and English. In contrast to earlier works that performed classification on the level of text lines or documents, we apply our method to the level of textual components, as we must first identify Chinese components before merging them into intact characters and sending the latter characters to a Chinese recognizer. To cope with a large training data set containing 365,672 samples, we employ a decision-tree support vector machine (DTSVM) method, which decomposes a given data space into small regions and trains local SVMs on those regions. By applying this method to train classifiers on various combinations of feature types, we were able to complete each training process within 3,500 seconds and achieve higher than 99.6% test accuracy in classifying a textual component into Chinese, alphanumeric, and punctuation. Moreover, the classification had no strong bias towards any of the three categories.

The attractiveness of human faces can be predicted with a high degree of accuracy if we represent the faces as feature vectors and compute their relative distances from two prototypes: the average of attractive faces and the average of unattractive faces. Moreover, the degree of attractiveness, defined in terms of the relative distance, exhibits a high degree of correlation with the average rating scores given by human assessors. These findings motivate a bi-prototype theory that relates facial attractiveness to the averages of attractive and unattractive faces rather than the average of all faces, as previously hypothesized by some researchers.

In this paper, we propose a number of adaptive prototype learning (APL) algorithms. They employ the same algorithmic scheme to determine the number and location of prototypes, but differ in the use of samples or the weighted averages of samples as prototypes, and also in the assumption of distance measures. To understand these algorithms from a theoretical viewpoint, we address their convergence properties, as well as their consistency under certain conditions. We also present a soft version of APL, in which a non-zero training error is allowed in order to enhance the generalization power of the resultant classifier. Applying the proposed algorithms to twelve UCI benchmark data sets, we demonstrate that they outperform many instance-based learning algorithms, the k-nearest neighbor rule, and support vector machines in terms of average test accuracy.

In this paper, we propose a new approach for identifying the language type of character images. We do this by classifying individual character images to determine the language boundaries in multilingual documents. Two effective methods are considered for this purpose: the prototype classification method and support vector machines (SVM). Due to the large size of our training data set, we further propose a technique to speed up the training process for both methods. Applying the two methods to classifying characters into Chinese, English, and Japanese (including Hiragana and Katakana) has produced very accurate and comparable test results.

A common practice in pattern recognition is to classify an unknown object by matching its feature vector with all the feature vectors stored in a database. When the number of class types and the set of stored entities are both large, it is essential to speed up the matching process. This can be achieved more effectively by eliminating unlikely candidates, rather than impossible candidates, from the process. We propose two methods for this purpose: one employs multiple trees, while the other employs a sub-vector matching technique. Both approaches use a learning procedure to estimate the optimal value of certain parameters. An online matching is then performed with a combination of the two methods, which matches candidates rapidly without sacrificing accuracy rates. The process is demonstrated by experiments in which we apply the proposed methods to handwriting recognition.

In this article, we propose a special type of decision tree, called a decision cascade, for binarizing document images. Such images are produced by cameras, resulting in varying degrees of brightness over different parts of the images. Our method decides what action to take on each part of the input image in order to obtain a satisfactory binary result. The advantage of this approach lies in its ability to learn decision rules from training data that may be labeled with multiple identities. Tests on images produced under improperly illuminated conditions show that our method yields much better visual quality and OCR performance than using a global threshold for binarization.

A new method for finding connected components frombinary images is presented in this article. The main stepof this method is to use a contour tracing technique todetect component contours and also to fill in interior areas.All the component points are traced by this algorithmin a single pass and are assigned either a new label orthe same label as their neighboring pixels. Experimentingon various types of document images (characters, pictures,newspapers, etc.), we find that our method outperformsthe other sequential methods using the equivalencetechnique. Our algorithm, moreover, is a method that notonly labels components but also extracts component contoursat the same time, which proves to be more usefulthan those algorithms that only label components.

We propose a new method that combines the use of a global threshold and a window-based scheme for computing local thresholds. The latter scheme compares the contrast of gray values within a neighborhood whose size varies with the scale of the objects being examined. To compute the scale quantity, a new wavelet model entitled Hadamard multiresolution analysis is also proposed. When the window-based scheme is applied to the areas where the global threshold is likely to fail, we obtain uniformly better binary results than using a global threshold only. Significant improvements to OCR performance can also be achieved by our binary results.

We propose a new thinning algorithm based on line sweep operation. Assuming that the contour of the figure to be thinned has been approximated by polygons, the "events" are then the vertices of the polygons, and the line sweep algorithm searches for pairs of edges lying within each slab. The pairing of edges is useful for detecting both regular and intersection regions. The regular regions can be found at the sites where pairings between edges exist. Intersection regions are those where such relations would cease to exist. A salient feature of our approach is that it finds simultaneously the set of regular regions that attach to the same intersection region. Such a set is thus called an intersection set. The output of our algorithm consists of skeletons as well as intersection sets, both can be used as features for subsequent character recognition. Moreover, the line sweep thinning algorithm is efficient in computation as compared with a pixel-based thinning algorithm which outputs skeletons only.

This article describes our research results in constructing a system that analyzes and recognizes printed Chinese documents. Emphasis is placed upon the interactions between different functional modules of our system that are necessary and important to the overall system performance. In particular, the interplay between image and document processing is seen as useful for removing spurious black spots in the background and white holes within character interiors. We also show that the interplay between feature analysis and character recognition enhances the generalization and discrimination power in recognizing characters. Our system holds above 98% recognition rate products. In some extreme cases, 25% better recognition rate has been seen. Our system is also able to treat many cases impossible for other systems.

In a previous article (Proc. 3rd Int. Conf. Document Anal. and Recogn., Montreal, Canada, pp. 227-30, 1995), we showed that a line sweep algorithm is an efficient means of line thinning. A line sweep is a process that works on polygonal figures and pairs the edges that bound the figure interior from two sides. In this article, we improve and extend this approach in the following way. First, a new method is used for grouping paired edges into regular and intersection regions. The regular regions can be found at the site where pairings between edges exist. Intersection regions, on the other hand, are where such relations cease to exist, due to the fact that pair relations between edges of wide distance are cancelled. Secondly, a salient feature of our new approach is to simultaneously find the set of regular regions that attach to the same intersection region. Such a set is called an intersection set. The output of our algorithm consists of skeletons as well as intersection sets. Both of them can be used as features for subsequent character recognition. Moreover, the line sweep thinning algorithm is efficient in computation as compared with a pixel-based thinning algorithm which outputs skeletons only.

We propose a pattern matching method for machine recognition of Chinese characters that are extracted from low quality document images. This method employs deformation representation, called character deformation model, to guide the detailed matching between templates and unknown characters. Experimental results show that the proposed method is very effective at recognizing deformed characters extracted from low resolution images or Xerox copy documents. A recognition rate of 95.7% is achieved for 200dpi document images, and the accumulated recognition rate of the first 3 candidates is higher than 99.8%. Moreover, even for the 100dpi document images, the third accumulated recognition rate is still above 91%.

Although there are thousands of many commonly used Chinese characters, they are actually composed of a lower number of stroke sub-patterns. Thus, in the task of recognizing optical Chinese characters, it is worthwhile to first identify the sub-patterns and then the whole characters composed of them. In such a hierarchical approach, however, decision mistakes at lower-levels can easily propagate into upper-levels to cause a high mis-recognition rate. To remedy this problem we devise a method called hierarchical flexible matching (HFM). The idea is to minimize the decision burdens at lower levels by allowing possibly conflicting sub-patterns to be identified from the same pool of primitives. The collection of these sub-patterns is then matched against some pre-specified models. In doing so, a metric is used to measure how well a candidate model is mapped into the given collection and how many primitives are covered by this mapping. We apply the HFM method to the font-indepdendent recognition of printed Chinese characters and have acquired very promising results.

We propose a new thinning algorithm based on line sweep procedures. A line sweep is a process where the plane is divided into parallel slabs by lines passing through certain "events" and then items are processed according to an order of the slabs. Assuming that the contours of the object that are to be thinned have been approximated by polygons, the "events" are then the vertices of the polygons and the line sweep algorithm looks for pairs of polygon sides that lie within each slab. Since the procedure is applied in both horizontal and vertical direction, possible conflicts may exist among the pairs of polygon sides. A subsequent effort is to resolve the conflicts according to a few generic types into which they can be classified. After the conflict resolution, the object can be decomposed into the regions that can be represented by the skeletons computed from the pairs of polygon sides, and the regions that are the singular parts of the object. Both types of regions can be used as features for subsequent pattern recognition operations.

A new stroke segmentation technique is reported. It is based on two essential operations applied to a given character: grouping adjacent segments into blade-like objects according to change of width as measured from global directions, and transforming those objects according to width variation as measured from intrinsic orientations. This technique proves to be effective in dividing Chinese characters into overlapping and non-overlapping components. Strokes are then restored, with the help of templates, by linking certain non-overlapping components terminating at the same overlapping areas. The resulting framework also lays down a basis for the structural matching of Chinese characters.>

The authors present traffic restoration design algorithms to attain a robust network for any facility link or node failure. Some of these algorithms were implemented on network models to compute the cost for doing traffic restoration for different traffic restoration level objectives. From their simulation studies, the authors found that with efficient use of trunk group diversity and at a reasonable incremental cost, one can obtain a robust network to respond to a major facility link or a node failure and still meet the network objectives. A similar approach based on the algorithms presented in this work has been proposed for design of the worldwide intelligent network.>