Shinji Tsuruoka

Modified Quadratic Discriminant Functions and the Application to Chinese Character Recognition

Issues in the quadratic discriminant functions (QDF) are discussed and two types of modified quadratic disriminant functions (MQDF1, MQDF2) which are less sensitive to the estimation error of the covariance matrices are proposed. The MQDF1 is a function which employs a kind of a (pseudo) Bayesian estimate of the covariance matrix instead of the maximum likelihood estimate ordinarily used in the QDF. The MQDF2 is a variation of the MQDF1 to save the required computation time and storage. Two discriminant functions were applied to Chinese character recognition to evaluate their effectiveness, and remarkable improvement was observed in their performance.

Improvement of handwritten Japanese character recognition using weighted direction code histogram

Several algorithms for preprocessing, feature extraction, pre-classification, and main classification are experimentally compared to improve the recognition accuracy of handwritten Japanese character recognition. The compared algorithms are three types of nonlinear normalization for the preprocessing, the discriminant analysis and the principal component analysis for the feature extraction, the minimum distance classifiers and the linear classifier for the high-speed pre-classification, and modified Bayes classifier and subspace method for the robust main classification. The performance of the recognition algorithm is fully tested using the ETL9B character database. The recognition accuracy of 99.15% at the recognition speed of eight characters per second is achieved. This accuracy is the best one ever reported for the database.

Increasing the feature size in handwritten numeral recognition to improve accuracy

The relationship between the recognition rate of handwritten numerals and the normality of the distribution of their features has been investigated experimentally with a large amount of data in various combinations of quantized orientations and regions. The recognition method is based on the histogram of local orientation of contours of each numeral. To obtain a more accurate orientation quantization, the effectiveness of the orientation quantization using the gray-scale gradient has also been investigated. The results show that : (1) to increase the dimensionality of features, it is better to increase the number of quantized orientations, keeping the number of regions small (e.g., 4 x 4 or 5 x 5) ; (2) in the same dimensionality, the better the normality of a feature distribution, the higher the recognition rate ; (3) a quantization of orientations using gray scales is effective for normalizing a feature distribution ; and (4) the filter processing in reduction of the number of quantization scales improves the normality and recognition rate. The recognition of handwritten numerals collected from actual posts were carried out by using the gray-scale local-orientation histogram (400 dimensions). A correct recognition rate of 99.18 percent (mean value) has been obtained.

Handwritten numeral recognition using autoassociative neural networks

Describes the result of a fundamental study on pattern recognition using autoassociative neural networks, and experimental comparison on handwritten numeral recognition by conventional multi-layered neural network and statistical classification techniques. As the statistical classification techniques, the projection distance method and the nearest neighbor method are employed. The relationship between the projection distance method which is based on the K-L expansion and three layered autoassociative networks is discussed, and it is shown that the three and five layered autoassociative networks are superior to the projection distance method. In the handwritten numeral recognition experiment, a total of 44862 numeral samples collected by IPTP are used to evaluate and compare the recognition rates of the autoassociative networks, the mutual associative network, the nearest neighbor method, and the projection distance method. The five layered autoassociative networks achieved the highest recognition rate in the handwritten numeral recognition experiment. The result of experiment together with the fundamental study show that the autoassociative networks have such characteristics that: (1) class independent training makes the possibility of local convergence less than that of the mutual associative network, (2) the networks possess the higher ability of dimension reduction and interpolation than the nearest neighbor method (3) they yield less misclassification due to subspace sharing than the projection method, (4) the five layered autoassociative network can fit a curved hypersurface to a distribution of patterns.

Automatic left ventricular endocardium detection in echocardiograms based on ternary thresholding method

This study proposes a new automatic detection method based on ternary thresholding method for echocardiograms. Two thresholds are determined by the discriminant analysis for the gray level histogram so that the input image is segmented into three regions: cardiac cavity, near epicardium, and the rest. Then the input echocardiogram is binarized with the lower threshold (between black and gray) to detect the cardiac cavity. The binary images are contracted n times to remove small regions and to disconnect the region of cardiac cavity from the other false regions. Among the obtained regions which corresponds to the cardiac cavity is selected and dilated 2n times to create a mask which restricts the region of the second thresholding operation. The masked image of each frame is binarized with another threshold determined by the discriminant analysis in the restricted area. Results of the evaluation test showed that the accuracy of the extracted contours was favorably compared with the accuracy of manually traced contours.The purpose of this edge detection and segmentation method for two-dimensional echocardiogram is to present the procedures to detect and segment an image from Two-dimensional echocardiogram and to generate a scanline that can be used to detect the distance between two endocardiums which is useful to analyze heart disease. This method applies image processing and computer graphic algorithms which were divided into 3 steps. Firstly, we used image improvement algorithms of noise suppression, histogram, brightness adjustment, threshold and median filtering. Then, edge detection algorithm with sobel compass gradient mask was applied to show the edge of endocardium border. Finally, segmentation and some computer graphics algorithms were used to identify and generate contour line of the endocardium border. Later in the study, Pearson correlation coefficient was used to evaluate performance of this method compared with that of manual track. The average correlation computes from this method is 0.9 which shows a good result because 0.9 is very close to 1. However, some part of contour line has a big error value. The unexpected result from incomplete of endocardium border came from color value of some part of border very close to background or noise color value. This problem occurred in first step can be solved by carefully collecting in collection process.

A segmentation method for touching Japanese handwritten characters based on connecting condition of lines

In unconstrained Japanese handwritten character strings, there are many touching characters. Segmentation of touching characters is required as preprocessing of isolated character recognition. But conventional segmentation methods cannot segment complicated touching characters. In this paper we propose a new segmentation method based on connecting condition of lines at a touching point, and evaluate the efficiency of this method for touching Japanese handwritten characters. This method could segment complicated touching characters with less unnecessary segmentations.

Obstacle to training SpikeProp networks — Cause of surges in training process —

In this paper, we discuss an obstacle to training in SpikeProp[1], which is a type of supervised learning algorithms for spiking neural networks. In the original publication of SpikeProp, weights with mixed signs are suspected to cause failures of training. We pointed out the cause of it through some experiments. Weights with mixed signs make the dynamics of the units activity twisted, and the twisted dynamics break the assumption that SpikeProp algorithm is based on. Therefore, it causes surges in training processes. They would mean an underlying problem on training processes.

Video Scene Segmentation Using the State Recognition of Blackboard for Blended Learning

We are developing the automatic generating system for video contents with keyword tag in blended learning. The generated lecture movie is too long for students to watch the lecture video on demand. We are considering the segmentation method of the lecture movie by the behavior of a lecturer. In this paper, we present a new segmentation method for one lecture movie to some shots using the behavior of a lecturer such as up-down movement of a blackboard and the erasure of characters on the blackboard. These behaviors are the sign of contents segmentation from the lecturer to students. The system detects the event time of these behaviors, and segments the lecture movie using the event time. We implemented our procedure and evaluated the validity using the lecture movie of some real lectures, and confirm that an accurate segmentation rate is 97%.

Region segmentation for table image with unknown complex structure

In this paper, we describe a system of region segmentation and conversion into an HTML file for an unknown machine-printed table image. Ruled lines delimit some cells of the table, and omitted ruled lines also delimit other cells. We consider a table analysis system for both types of table cell. First, our system segments a table by means of the ruled lines into some regions. Secondly, these segmented regions are further segmented into cells by the omitted ruled lines that are indicators (such as numerals and characters). The cells include several character lines, and our system can convert a table of unknown complex structure into an HTML file. Also, we confirm the effectiveness of our region segmentation method for various kinds of tables with omitted ruled lines by computer experiments.

Extraction Method of Retinal Border Lines in Optical Coherence Tomography Image by Using Dynamic Contour Model

In the field of ophthalmology, the needs of retina diagnosis using optical coherence tomography (OCT) images have been growing, and the automatic measurement of a retina thickness and its quantitative evaluation are desired for the diagnosis of retinal diseases. Previously, the automatic measurement methods of the retinal thickness have been reported for retinal OCT images. These previous methods can extract the retinal border lines (ILM and RPE) appropriately in most cases of normal OCT image. However these methods caused the tracking error to some OCT images with large noises. In this paper, we propose a new automatic measurement method of a retinal thickness in OCT image. The method employs ODAN (One Directional Active Net) to extract ILM and RPE. ODAN employs a new energy function to extract the retinal border lines exactly and all nodes of ODAN moves only to one direction to minimize the total energy repeatedly. The energy function consists of (1) the conformity characteristics energy of image and (2) the internal strain energy. We confirmed the usefulness of the ODAN by the experimental results for ten OCT images with large noises. We compared the positions of retinal border lines by the proposed method with the positions in a manual trace by ophthalmology specialist. In the comparative result, the proposed method is useful as the basic method for the detection of retinal diseases.