Scale-space image theory, in which a Gaussian filter with set scale parameters is used to generate a series of blurred images, has become the basic technique of computer vision. We have proposed to apply a spectral theory to compress the scale space and demonstrated some applications such as blur image generation and SIFT. In this paper, as another application of spectral theory, we propose template matching method for blurred image. A basic idea of the proposed method is changing the scale parameter of the template to fit the input blurred image. Our experimental results show that the proposed method detects the XY position and scale of the target object accurately.

View full abstract

We propose a hand-waving detection method robust to similar behaviors and illumination change. The system detects moving regions in the image using optical flows and judges whether the region is corresponding to hand-waving or not based on four metrics : distribution shape of optical flows, motion in the depth direction, moving direction in the four regions around the hand, and the number of voting for iterative motions. Experiment shows that false positive rate is less than 10% and the detection rate is more than 90%.

View full abstract

This paper describes an active stereo method that works with interreflection on specular surfaces. We use two cameras and a projector to take two series of images for depth maps with active stereo systems. Comparing the depth maps, we judge the validity of measured depth. For the valid depth areas, we turn off corresponding pixels of the projector and again take two series of images for depth maps that suffer from less interreflection. Iterating depth estimation and projected-pattern extinctions, we could mitigate interreflection and increase measurable areas for 3D reconstruction. Experiments show the improved results of the proposed method for specular surfaces with strong interreflections.

View full abstract

The purpose of this research is to develop automatic raw shiitake grading system using computer vision technique. This paper shows the image analysis method to judge the level of opening cap. There are raw shiitake mushrooms on blue background in the material images. The upper side of these raw shiitake mushrooms is lamella. Proposed method uses the lamella area of shiitake mushroom because the color of raw shiitake mushroom cap is almost same to lamella area. The gradient vectors in lamella tend to be perpendicular in the vectors from the center of shiitake area. The lamella area is extracted by this tangency. The rate of lamella area is used to judge the level of opening cap. If raw shiitake mushroom has stem, the rate of lamella area in half area that doesn't include stem area is used. AdaBoost algorithm is used to finally judge. Experiment shows that the rate of learning data is 92%.

View full abstract

There are many such practical cases that it is difficult to detect defects even by a visual inspection. These defects, such as very small cracks of pattern and thin stains, etc., are slightly different with each other and hidden in the background texture. Even when the skilled inspector can detect these defects, they are undetectable by the traditional image processing algorithms. This paper describes the detailed program for detecting irregular region involving such invisible defects inspired by human vision architectures. This program has been named “KIZUKI” algorithm in this study. Two kinds of typical vision theories are used to model the “KIZUKI” process : the peripheral vision and the involuntary eye micro-vibration theories found in the field of psychology of vision. It consists of three steps : first step for forming low resolution images by a variable re-sampling, and the second step for detecting relatively salient regions at each low resolution image, and the last step for integrating these regions. Experimental results indicate that our approach is very reasonable in intuitive interpretation of human vision. Since our programs are very simple in architecture and effective in performance, it might be applicable equally to the quite different kinds of defects.

View full abstract

We propose a system for pose estimation of an object and misestimation detection. The system estimates object pose by exploring a three-layered search tree whose node is a model image of appearance. The model images are generated from different points of view using a 3D object model in advance. Matching between model and object images is performed using SIFT feature and RANSAC. However, the pose estimation sometimes fails, especially due to occlusion. Hence, the system evaluates the result of pose estimation by shape fitness using weighted voxels. Experimental results show the basic validity of the system. The system detects misestimation even if the object is partially occluded by other objects.

View full abstract

This paper introduces a reliable 3-D object recognition method which can statistically minimize its mismatching. This method basically uses a 3-D object model which is represented as a set of feature points with 3-D coordinate. Each feature point also has an attribute value about local shape around the point. The attribute value is represented as an orientation histogram of normal vector, which can be calculated by using several neighboring feature points around each point. Here, this attribute value means its local shape. By estimating relative similarity of two points of all possible combination in the model, the distinctiveness of each point is defined. In the proposed method, only a small number of distinctive feature points are selected and used for matching with all feature points extracted from an acquired range image. Finally, position and pose of the target object can be estimated from some correctly matched points. Experimental results using real scene have proved that recognition rate of this method is 93.8%, which is 42.2% higher than the conventional Spin Image method. Also, computing time is about nine times faster than that.

View full abstract

In this paper, a novel image feature named Sparse FIND which is suitable for onboard pedestrian sensor due to having low computational cost is introduced. Histograms of Oriented Gradients (HOG) have been shown excellent capability of image classification. Feature correlation descriptor (FIND) which has achieved higher capability of image classification in comparison with HOG has been developed. However, FIND requires much more computational cost than HOG due to computing the correlation between histogram elements. In Sparse FIND, the computational cost is reduced by limiting the computation of the correlations between histogram elements which contribute classification majorly. Through experiments, Sparse FIND performs nearly the same capability of image classification as FIND does despite the computational cost is approximately one tenth in comparison with it of FIND, is confirmed.

View full abstract

Super-resolution is a technique to estimate a high-resolution image from plural low-resolution images and has been applied in various fields. In the semiconductor device manufacturing there has been a pressing need for inspection and measurement by using high-resolution image of scanning electron microscope (SEM). Super-resolution is expected to be effective, while it may fail in estimation of high-resolution image due to the brightness fluctuation between corresponding pixels in low-resolution images. We propose a novel super-resolution method that can handle plural factors of the brightness fluctuation based on a maximum a posterior probability (MAP) estimation. The proposed method maximizes a posterior probability consisting of combinations of plural fluctuation probability distributions for shot noise, pattern deformation, and exception value (e.g., defect particle). Robustness of the proposed method against these fluctuation factors is shown by using real SEM images of semiconductor patterns.

View full abstract

Our group is developing prenatal DNA diagnosis with Nucleated Red Blood Cells (NRBC) of fetuses extracted from blood of their mothers. In this diagnosis, it is important to detect NRBC automatically from microscope images for efficient examining. In a research, an automatic NRBC detection system is proposed with two parts. The first part extracts candidate nucleuses with rule-based algorithm. The second part discriminates cell regions with machine learning algorithm. In the second part, Cell-HOG feature specialized for detecting NRBC is developed. Cell-HOG feature is customized HOG feature that emphasizes differences between NRBC and other cells such as white blood cells, out of focus cells, and red blood cells with foreign objects. By using Cell-HOG feature, number of false positives becomes to smaller than that of HOG feature.

View full abstract

Conventional transparent film thickness measurement methods such as spectroscopy are essentially capable of measuring only a single point at a time, and their spatial resolution is limited. We have developed a wide-view film thickness measurement system based on an extended global model-fitting algorithm developed for three-wavelength interferometric surface profiling. It estimates the film thickness distribution from an interference color image captured by a color camera with three-wavelength illumination. The basic performance was experimentally proved, and the system was successfully applied to measure the air gap thickness of Newton's rings and the thickness of a vertical flowing soap film. Its key features are : (1) high spatial resolution, (2) high-speed measurement, (3) low cost and simple optics, and (4) no preliminary calibration required.

View full abstract

A skin detection method by using the reflectance characteristics of the human skin in near-infrared ray band has been proposed. This method is one of the robust method to detect the human skin region in real time. However, the human skin may not be detected correctly when the light at unknown spectral ratio of three wavelengths affects it. In order to solve this problem, we introduced IR3 camera which can take three images of near-infrared wavelength, (870nm, 970nm and 1050nm) at the same time. In addition, we used a reference light source at known spectral ratio of three wavelengths. In our proposed method, the effect of the unknown ambient light is suppressed by subtracting images which are irradiated and not irradiated with the reference light source. However the problem with, this method is that the influence of the unknown ambient light is not suppressed under the sunlight. Because the sunlight is a very strong light, and the spectral ratio of the sunlight changes by the minute. In this paper, we divided this problem into two cases to solve this; shadow regions and sunlight irradiated regions. In the case of the sunlight irradiated region, the human skin region is detected by estimating the spectral ratio of the sunlight to equalize the balance of the irradiation of three wavelengths in the sunlight. In the case of shadow region, the human skin is detected by suppressing method of the unknown ambient light. By merging these two results, all of the skin regions are detected under the sunlight.

View full abstract

This paper presents comparative evaluation of a robust frame-level descriptor for video identification, called a Video Signature, which was previously developed by the authors and was accepted as part of a new ISO/IEC standard “MPEG-7 Video Signature Tools”. The Video Signature is designed for video copy detection from a large-scale database, with robustness to various modifications, including caption overlay, camera capturing, compression, and color changes, working at extremely low false positive rate. It represents intensity differences between various sub-regions in a frame, which are configured at variety of scales, shapes, and locations in the frame, to provide robustness and uniqueness to the descriptor. The intensity differences are quantized into ternary values, resulting in a compact representation of 76 bytes per frame. Furthermore, a confidence value is calculated for each frame, which is used in the matching process to significantly reduce false positives. An experiment to identify videos under nine types of modifications was carried out to compare the performance of the Video Signature to that of the conventional color-based and spatial visual features. The results show that the Video Signatures improves the detection rate for all modification types compared with the conventional features, in particular significant increase by 38.8% for caption overlay and 61.8% for camera capturing have been achieved. We also present a system for structuring a large-scale video database using the Video Signature, developed for efficient management of video contents. The details of the system are explained and its feasibility as an application is shown.

View full abstract

A novel method that evaluates white uniformity grade of electronic displays based on visual characteristics is presented. The original color images are converted into opponent color images. The images are transformed into frequency domain, and filtered by the color contrast sensitivity functions. The mura (non-uniform region) index E is defined to evaluate the white uniformity grade quantitatively. The 168-pseudo samples which color saturations,mura areas and edge gradients values are different, are prepared. The experimental results showed strong relations between human sensory evaluation scores and the mura index E values for the 168 samples, and the validity of the proposed method was demonstrated.

View full abstract

In this paper, a method to reconstruct 3D trajectory of a table tennis ball and two systems that obtain shot information for table tennis performance analysis are proposed. The method and the systems are developed to solve the problem of conventional performance analysis in table tennis. There are several methods to reconstruct 2D ball trajectory or 3D trajectory of the heavy ball compares to a table tennis ball. However, the methods cannot be adopted for 3D trajectory reconstruction of a table tennis ball, because there are problems attribute to the dimension of a trajectory and the weight of a ball. The proposed method in this paper can reconstruct 3D trajectory of a table tennis ball. The key feature of the method is to approximate that a ball travels on tilted planes. The approximation makes the reconstruction robust against failure to measure 3D ball positions in some frames. Based on the proposed method, a system using two RGB cameras is developed. The system is experimentally verified that it can provide accurate information for match analysis. Next, a system using an RGB-D camera is developed to optimize usability for practitioners. The system is experimentally verified that it can provide accurate information for service analysis.

View full abstract

This paper proposes a high-speed and high-accuracy matching method for 3-D dot cloud data, and a method of expanding the scope of the matching. The proposed algorithm is designed to reduce the matching cost of ICP (Iterative Closest Point) algorithm by reducing the number of data. When the number of the data is decreased, the accuracy of the matching tends to become lower. However, we change the dot cloud data varying in the reduction rate progressively as the matching process progresses, and perform the matching with high speed and high accuracy. In this research we propose two types of stage processing depending on the magnitude relationship between the number of online data and offline data. One is the movement point type where online data is larger than offline data, and the other is the reference point type. Moreover, we expand the scope of matching by changing the number of reference points for the search points. We name this k-nearest neighbor stage processing. Our proposed methods are combined in order to achieve high-speed and large scope matching, which result in two techniques; the movement point type k-nearest neighbor stage processing, and the reference point type k-nearest neighbor stage processing. We perform some experiments using real data measured by a laser ranger to verify the effectiveness of the proposed techniques.

View full abstract

Recently, many studies for autonomous mobile robot are being conducted for applications in outdoor environment and Unmanned Ground Vehicle (UGV) in particular is getting much attention. Most of those studies focus on environment where an autonomous vehicle can separate navigable area from non-navigable area clearly. Therefore, conventional approaches can have troubles in classification in off-road environment, where navigable areas cannot be obviously identified. To overcome this problem, we propose an approach which enables a UGV to autonomously identify where it can safely steer for. First, we propose to use linear Support Vector Machine for terrain classification in off-road by extracting multiple features. Second, we propose an implementation of potential-based method for extracting traversable direction. The performance of the proposed approach was verified in an off-road environment and it was shown that the proposed method is superior to a conventional approach.

View full abstract

This paper proposes a method of detecting affected segments of glaucoma from optical coherence tomography (OCT) images. Thickness of nerve fiber layers and its asymmetry, difference, and variance are used as features. OCT images are segmented and the four features are obtained at each segment. Normal and glaucoma classes are constructed at each segment using training data. Detection of affected segments of glaucoma is carried out using four pattern classification methods : classification using Mahalanobis distance, maximum likelihood, nearest neighbor method, and support vector machine. The proposed method is evaluated by experiments to compare the detection results by the method and ones by a doctor.

View full abstract

This paper proposes a rotational registration algorithm for point sets that represent defects on the surfaces of hard disks or other disk media that lack registration indexes. The proposed algorithm is designed to match point sets more accurately. It works by using an existing matching algorithm called Local Consistency of Point Dispersion (LCPD) to estimate the numerical ranges containing the true rotation angle while adjusting the point set's rotational position, enabling the algorithm to iteratively adjust the LCPD parameter values.

View full abstract

This paper presents an extended shadowing compensation model for eliminating shadowing errors in the scanning electron microscope photometric stereo method. This extended model removes the earlier restrictive assumption that the shadowing effects exist only on horizontal surfaces. This model is developed by modeling image intensities in both the absence and presence of shadowing. A numerical solution is proposed to resolve the implicit expressions for double integrals in the model. Compared to the previous shadowing compensation model, the extended model can be more accurate when applied to shadowing regions on surfaces having large inclinations.

View full abstract

In recent years, researches on the habit of creature and its characteristic are widely treated based on the image observation. Especially, research of creature which transmits a pathogen is important. For instance, at Japan environmental sanitation center, evaluation experiment of medicine evasion to cockroach is conducted. Then, since the evaluation is done by human-eye observation, both of time cost and misidentification by human are problems. Then, in order to improve these problems, the method of creature observation which uses the image processing is one of key roles. Namely, development of method which automatically derives the action trajectory of creature and does creature observation quantitatively is important. In this paper, we introduce new method of detection of cockroach based on advanced statistical reach feature, and generate the action trajectory of cockroach. Further, through experiments we prove the effectiveness of our proposed method.

View full abstract

In this paper, we propose a marker-less motion capture method for a human face. We detect the movement of a face as the deformation from the initial pose to other poses. In such cases, the non-rigid registration algorithms are commonly used to detect the movement of an object. However, since the features of the movement of a face vary in each facial part, it complicates the registration problem. In our method, we recognize the facial parts by using Random Forest algorithm to be simplified to solve the registration problem. Then, we classify the facial parts into five types : nose, mouth, eye, cheek and obstacle. And we define the feature vector of each point of the face by using Fast Point Feature Histograms and a normalized position. We train the Random Forest based on the feature vector to recognize the facial parts from an arbitrary point set data. After the recognition of the facial parts, we detect the movement of the each part by using a non-rigid registration algorithm. Finally, we combine the translations of the each point as the deformation of the face by using the deformation technique based on a Radial Basis Function. In our results, we show that the proposed method enable us to detect the motion of the face more accurately.

View full abstract

Recently, the number of emergency vehicles equipped with light-emitting diode (LED) warning lights has increased. Unlike traditional beanie lights, LED warning lights can display various flickering patterns because these patterns are controlled by computers. The purpose of this study was to develop flickering patterns that had a high level of visibility. Lighting time (ON time), no-lighting time (OFF time), light intensity and rising time (UP time) were controlled by a microcomputer. Further, the visibility of each LED flickering pattern was evaluated quantitatively. Specifically, the visibility of the flickering patterns using the psychophysical measure of “conspicuity (easy to stand out)” from Bradley-Terry paired comparison model was identified. Results showed that OFF time had a significantly greater influence on visibility than ON time. Moreover, the flickering pattern with 66 msec of OFF time provided optimal visibility, regardless of the ON time, by making the visibility map, which represented the degree of visibility between ON time and OFF time. Therefore, the ideal combination between ON time and OFF time was determined.

View full abstract

The need for rendering woven fabrics arises frequently in computer graphics. Woven fabrics have a specific appearance, sheen, and transparency. A BRDF (bidirectional reflectance distribution function)/BTDF(bidirectional transmittance distribution function) model is well known as the basic technology for expressing the appearance of woven fabric. In order to represent the transparency of woven fabric, a BSDF (bidirectional scattering distribution function) model is required in addition to the BRDF/BTDF model. In this paper, we propose a BSDF model using the microfacet BTDF model(GGX model), and perform comparison experiments. As a result, we confirmed the effectiveness of our method with objects of quality of generated CG image, rendering time, and parameter adjustment. Our goal is to produce a catalog of curtain animations that can express various kinds of Woven Cloths under arbitrary light conditions.

View full abstract