Satoru Igarashi

Robust image registration by increment sign correlation

Abstract A novel and robust statistic as a similarity measure for robust image registration is proposed. The statistic is named as increment sign correlation because it is based on the average evaluation of incremental tendency of brightness in adjacent pixels. It is formalized to be a binary distribution or a Gaussian distribution for a large image size through statistical analysis and modeling. By utilizing the proposed statistical model, for example, we can theoretically determine a reasonable value of threshold for verification of matching. This sign correlation can also be proved to expectedly have the constant value 0.5 for any uncorrelated images to a template image, and then the property of the constancy can be utilized to analyze the high robustness for occlusion. The good performance for the case of saturation or highlight can also be proved through theoretical analysis and fundamental experiments. A basic algorithm for image scanning, search and registration over a large scene is represented with a technique for a fast version by the branch-and-bound approach. Many experimental evidences with real images are provided and discussed.

Using selective correlation coefficient for robust image registration

A new method is proposed for robust image registration named Selective Correlation Coefficient in order to search images under ill-conditioned illumination or partial occlusion. A correlation mask-image is generated for selecting pixels of an image before matching. The mask-image can be derived from a binary-coded increment sign-image defined from any object-image and the template. The mask-rate of occluded regions is theoretically expected to be 0.5, while unoccluded regions have much lower rate than 0.5. Robustness for ill-conditioned environment can be realized since inconsistent brightness of occluded regions can be omitted from the mask operation. Furthermore, the mask enhancement procedure is proposed to get more stable robustness. The effectiveness of masking increases by the procedure, resulting in the rate around 0.7 for masking of occluded regions. This paper includes a theoretical modeling and analysis of the proposed method and some experimental results with real images.

Reaction heat effects on initial linear shrinkage and deformation in stereolithography

In the industrial use of stereolithography, precision is always a problem. The basic phenomenon of solidification shrinkage has not been sufficiently investigated. This study aims at clarifying the initial linear shrinkage of cured resin in a minute volume. Experimental equipment has been developed which measures the time history of the single strand in situ in a stereolithography machine. An analysis model of the time history of a minute volume linear shrinkage was shown using the measured shrinkage of a cured line segment. The relation between the time history of the linear shrinkage and temperature was measured and the shrinkage in the minute volume after irradiation was found to result due to temperature variation. Deformation and linear shrinkage were measured with two scanning orders to control the thermal distribution in layer forming. The effects of thermal distribution were also observed in one layer forming.

Robust line fitting using LMedS clustering

The authors propose a robust clustering method based on the Least Median of Squares principle. This is a general-purpose method, and is efficient for finding a majority structure using dictionary sorting and a smallest square region enclosing for a set of data points. In addition, the authors propose a robust line fitting algorithm using this method. The fitting problem is solved through a dual transform of a pair of points selected from data space to a parameter space, and then clustering the mapped points. Moreover, a function to identify outliers based on the size of the square region converging in the parameter space is defined, and is used to distinguish between normal values and outliers. The proposed method enables robust line fitting for data points found in images or in files. The resulting line statistically satisfies the least median condition. Furthermore, line fitting to groups of points with a high outlier percentage is also possible. Here, outliers and normal values are removed in sequence. The validity of the proposed method has been shown based on the results of simulation experiments.

A Refined Theory of Anisotropic Thick Plates

For the deformation of anisotropic plates with arbitrary thickness, a refined theory is proposed to derive approximate equations with any desired accuracy, and general two-dimensional equations in the n-th order approximation are presented. The approximate equations presented are solved for a graphite/epoxy composite plate subjected to a sinusoidally distributed load at the upper surface, and the solutions are compared with the exact one and with those of several other approximate theories by numerical calculation. The results obtained are as follows: (1) The accuracy of solutions of equations in any given order of approximation decreases with increase in the strength of anisotropic property as well as in the plate-thickness. (2) As the order of approximation is increased, the accuracy of solutions is improved and the solutions approach the exact one asymptotically. (3) When the accuracy of solution is appointed for a plate with any thickness and elastic properties, the order of approximation of equations to be used can be determined by preliminary calculation as shown in this paper.

Object Recognition Based on 2-D Orthogonal Expansion of Image by Marginal Eigenvectors.

This paper proposes a new method to recognize 3-D objects and their poses using the approximate representation of an image based on the 2-D orthogonal expansion by marginal eigenvecters of column- and row- covariance matrices. The proposed method considers 2-D structure of images, and executes learning process by treating low dimensional matrices, so it can learn a number of images with small computation cost compared with the methods based on the K-L expansion. The method can also recognize any pose of an object, which is not learned, using smooth interpolation between highly correlated discrete images obtained in learning stage of the object. In this paper, the learning and descriminating procedures are formulated and effectiveness of the method is shown through fundamental experiments with real objects.

3D Measurement of Shape Using Differential Stereo Vision Alborithm. Enlargement of Depth Range and Improvement of Measurement Accuracy.

In usual 3D measurement of shape by using stereo image, it is necesary to solve a correspondence problem which requires rather complicated image processing and high computational cost. Recently Ando developed the differential stereo vision system in which depth information can be derived from brightness values of image without solving the exact correspondence problem. However, this method has a weak point of narrow depth range in measurement. In this paper, we propose a method to solve a narrow depth range problem and to improve the accuracy of measurement for this method. Several examples of measurement are presented to demonstrate the effectiveness of the proposed method.

Fundamental study on automatic detection and discrimination of surface defects in wood.

There are many types of defects in wood, some of which reduce strength of structure parts or spoil the appearance of furniture parts. To make efficient use of valuable wood free from such defects, it is desired to develop an automated inspection system which is able to locate the position and extent of each defect present and to identify the type of defect at each location. In this paper, we propose an automatic inspection method to detect and discriminate surface defects in wood. The essential points of the method are as follows : Input gray level image of testing wood is divided into a number of regions and several tonal property measures are extracted from image of each region, and fuzzy clustering technique is adopted for detection and discrimination of surface defects. Applying the proposed method to inspection tests of actual lumbers, we obtained the following results : (1) Defect detection can be achieved almost completely. (2) Discrimination rate ranges from 25 to 82% depending on the type of defect, and this indicates that more research is needed.