/ Shunji Mori / Professor
/ Tony Y. T. Chan / Assistant Professor
/ Yu Nakajima / Research Associate
First and foremost, the Image Processing Laboratory engages in research and development of image pattern recognition systems. More specifically, as can be seen from the background information and the recent research publications of the members of the laboratory, character recognition is our current focus. Related to the recent involvement of multimedia systems, character recognition has been noted by many researchers and engineers. On the other hand, character recognition techniques are generally divided into off-line and on-line methods. The former is typical in character recognition and aims at duplicating the human ability in recognition. However, on-line character recognition is also noted recently in connection with so-called pen computers. On-line methods provide very flexible, convenient, and natural human-interface. Historically speaking, these techniques have been developed separately. However, both techniques can be developed together in principle. The common approach makes possible recognition flexibility in such a way that the usual constraints being imposed on the on-line techniques can be removed. For example, writing order and number of stokes constituting a character are typical of such constraints. On the other hand, we are requested to develop an inspection system for flaw of Sake bottles. The contribution to the domestic industry is very important mission of our University. Therefore, we are doing a basic research on that together with Industry Application Laboratory. In general, image inspection is a very important theme and has many applications in industry broadly. Prof. Mori and Mr. Nakajima have engineering background. Actually, Prof.Mori was a leader of an OCR group when he worked for Ricoh Research and Development Center and developed a very powerful OCR package which is used in practice. One of the members of the OCR group was Mr. Nakajima, who has joined us as a research associate. He is well-acquainted with different computer systems and also works as an instructor in programming. The system developed is called filtering in character recognition, which makes the packages very fast without being hardware-specific. The OCR group were awarded the Excellence in Programming Prize from the Ministry of Post and Telecommunications. The basic configuration of the system was proposed by Prof. Mori.
Naturally, the essential point of the developed system lies in the excellent algorithm of character recognition, which is based on an algebraic approach at shape description. This approach was investigated as basic research by Prof. Mori when he worked for the Electro-technical Laboratory of the Ministry of International Trade and Industry. Later the basic line was further developed, in theory and in practice, by Prof. Nishida. We think that is one of the most elegant theories in character recognition. Because of its theoretical richness and high-level abstraction, the very broad applications in shape recognition in general. It is applicable to both off-line and on-line character recognition systems, for example. Further extension of the theory is in the application of the recognition of three-dimensional bodies, in which a piece of the surfaces must be represented effectively in terms of pattern recognitions and global organization mechanism. In this regard, a set of primitive surfaces must be considered. Furthermore, in addition to basic recognition mechanisms in a bottom-up manner. rich knowledge concerning "objects" must be naturally employed. Some knowledge representational methods developed in the field of artificial intelligence research are being considered. Assistant Professor Chan is investigating unifying pattern recognition and artificial intelligence approaches based on models proposed by Prof. Lev Goldfarb of the University of New Brunswick, Prof. Nishida, and himself.
Refereed Journal Papers
A general (non-heuristic) computational analytical model to tackle the difficult unsupervised inductive learning problem is proposed by making some additions and modifications to an existing metric model so that the model is more elegant and able to handle the unsupervised case. It turns out that it is instructive to treat, in essence, the supervised problem with noise as an unsupervised problem. We demonstrate the success of the new model on the benchmark XOR (exclusive-or) and parity problems by showing how the inductive agent successfully learns the weights in a dynamic manner that would allow it to distinguish between bit-strings of any length and unknown labels.
Basically, image processing deals with ways of preparing raw images so that they are suitable for human or for machine recognition. The two purposes often require very different methodologies. As a result, the field of image processing touches on many subject areas from electronics to the psychology of human perception. Russ approaches the field from an empirical engineer/practitioner perspective with an emphasis on pedagogy.
Refereed Proceeding Papers
Two methods for stroke segmentation from global point of view are presented comparatively. One is based on thinning method. The other is based on contour curve fitting. For both cases an input image is binarized. For the former, Hilditch's method is used. Then crossing points are sought, around which some domain is constructed. Outside the domain, a set of line segments are identified. These lines are connected and approximated by cubic B-spline curves. Smoothly connected lines are selected as segmented curves. This method works well for some limited class of crossing lines, which are experimentary shown. The other is that a contour line is approximated by cubic B-spline curve, along which curvature is measured. According to the extreme points of the curvature graph, the contour line is segmented, based on which the line segment is obtained. Experimental results are shown for some difficult cases.
A model is proposed for learning to classify patterns under the Euclidean setting. Each pattern is represented by a vector in a fixed D-dimensional Euclidean space. Patterns are divided into training and test sets. Eleven experiments were performed. The proposed naive learner is found to be extremely fast and yet the correct classification rates are respectable even when compared with some of the best known rates.
A simple approach to feature subset selection is proposed. During the training stage, the method selects the features that simultaneously minimize the within-class distance and maximize the between-class distance. Experiments performed on the Iris Plants Database and the Pima Indians Diabetes Database show that the approach is practical because it is fast and yet the correct classification rates are competitive.",
In this paper, we experimented on the line fitting using the multiple B-spline curve. The nature of the approximated B-spline curves in terms of curvature was examined. The approximated B-spline curve has some conditions; the number of degree, the number of segment, and the boundary condition. In order to the effect of these conditions to the approximated B-spline curves was investigated. Specifically, we generated two types of Bezier curves; gentle and sharp curves, each of which was approximated based on least mean square method. Thus, the original Bezier curves were compared with the approximated curves with respect to the curvature features. From the detailed research, we obtained the combination of conditions of approximated B-spline curve to the curve types. Also we applied the B-spline approximation method to a real data curve. To utilize our approximated method an extra part of the curve was cut and, we obtained the peak value of the curvature clearly.
This research is an attempt to quantify aesthetical features of Chinese character (Kanji) with the view of finding out some common factors of beautiful form, as a side research of character recognition. Assuming that printed characters can be generally accepted as "beautiful" characters, the features of printed Kanji are calculated. The feature indices used here are center of gravity, spread, and complexity based on the method of moment.
First, all the Kanji data are calculated, and it was found that they have some coherent distributions which seem to contribute the stability and unity of articles. Second, some Kanji groups are selected and each character is decomposed into its radical components. Typical feature formulae were applied to the components and to the whole characters, and common coherent results were obtained in each Kanji group. The tendency of taking balance of the whole character by adjusting components can be read from the value distributions. A fact was found from this calculation that even a common radical of fixed form is designed one by one according to the each character's structure. Third, the characters having extreme feature values were picked up from the Kanji groups and these characters of other fonts were caluculated in the same way. Center of gravity and spread show fairly consistent results through all the fonts though it was not so clear in complexity. These results indicate the common sense of taking balance among different fonts designers.
From the numerical features, it was found that "balancing" is one of the important factors of beauty of characters. Also it was found that the balancing process is common to different font designers though the fonts themselves show different appearances.
In this paper, we proposed a very effective rough classification scheme of on-line Kanji. In the case of on-line character input each stroke is segmented, so we used this advantage fully. That is, our scheme is based on stroke number, stroke shape and intersection of stroke. Here we assume that the writer always uses the same number of strokes when they write Kanji. Stroke shapes are approximated cubic B-spline function and basic features such as direction and curvature are easily calculated. For intersection, we developed a simple method to find intersecting points. The experiment for a Kanji set which consists of 936 classes was conducted. The result demonstrated that the classification scheme is very effective.
This paper presents a new method of extracting relational feature between strokes in on-line Kanji characters. In on-line input, each stroke in a Kanji character is easily segmented. Therefore, the next problems are types of each stroke and the relational feature between strokes. In this paper, we treated the latter based on Glucksman's method in background analysis. This is beginning stage of the research and some typical experiments are shown.
A method for segmentation is presented. A binalized image is used and thinned by Hilditch's method. Then, we examine the line connection at the place of crossing point. Near the neighbor of a crossing point, a black box is constructed and virtually the detailed structure is neglected. In the outside of the black box, regular line segments are connected as candidate segmental lines, which are approximated by cubic B-spline curves. Then smoothest candidate is selected. This scheme was implemented by Sato and attained some degree of success. However, there were still problems, which were improved in this research. This is demonstrated in this paper.
A database of an online hand-written character set of the education Kanji characters (996 characters) was made. The on-line characters were written according to the standard stroke writing order. We have also developed Kanji data-base and a retrieval interface so that the data-base can be easily used.
In off-line handwritten character recognition, identification with high accuracy has not been achieved yet despite much efforts. Problems for the improvement on accuracy is vague in correlation method in particular. We constructed an off-line handwritten character recognition system, and confirms problems through the computer simulation. In addition, analysis of recognition errors indicates the direction for the improvement.