Kota Aoki

Cross-Parameterization for Triangular Meshes with Semantic Features

We propose a technique for fusing a bracketed exposure sequence into a high quality image, without converting to HDR first. Skipping the physically-based HDR assembly step simplifies the acquisition pipeline. This avoids camera response curve calibration and is computationally efficient. It also allows for including flash images in the sequence. Our technique blends multiple exposures, guided by simple quality measures like saturation and contrast. This is done in a multiresolution fashion to account for the brightness variation in the sequence. The resulting image quality is comparable to existing tone mapping operators.In 3D computer graphics, mesh parameterization is a key technique for digital geometry processings(DGP) such as morphing, shape blending, texture transfer, re-meshing and so on. This paper proposes a novel approach for parameterizing a mesh into another one directly. The main idea of our method is to combine a competitive learning and a leastsquare mesh techniques. It is enough to give some semantic feature correspondences between target meshes, even if they are in different shapes or in different poses. We show the effectiveness of our approach by giving some examples of its applications.

Segmentation-based illumination normalization for face detection

Face detection is an important research topic in the field of computer vision. Illumination problem is one of the most important aspects impeding the effectiveness of face detection. The well known Haar-like face detector developed by Viola and Jones is also largely weakened under adverse lighting conditions such as backlighting or uneven lighting. In this paper, a novel segmentation-based illumination normalization method is presented for the purpose of compensating non-uniform illuminations and increasing the robustness of Haar-like face detector. First Otsu method is employed to segment the input image. Then the proposed illumination normalization method called Half Histogram Truncation and Stretching (HHTS) is applied to locally attenuate the illumination and enhance the visibility of local patterns (facial structures). Finally Haar-like face detector is executed to locate faces. Experimental results show that it can remove non-uniform illuminations efficiently and significantly increase the performance of the original Haar-like face detector.

3D Triangular Mesh Parameterization with Semantic Features Based on Competitive Learning Methods

In 3D computer graphics, mesh parameterization is a key technique for digital geometry processings such as morphing, shape blending, texture mapping, re-meshing and so on. Most of the previous approaches made use of an identical primitive domain to parameterize a mesh model. In recent works of mesh parameterization, more flexible and attractive methods that can create direct mappings between two meshes have been reported. These mappings are called “cross-parameterization” and typically preserve semantic feature correspondences between target meshes. This paper proposes a novel approach for parameterizing a mesh into another one directly. The main idea of our method is to combine a competitive learning and a least-square mesh techniques. It is enough to give some semantic feature correspondences between target meshes, even if they are in different shapes or in different poses.

Brain tumor classification of microscopy images using deep residual learning

The crisis rate of brain tumor is about one point four in ten thousands. In general, cytotechnologists take charge of cytologic diagnosis. However, the number of cytotechnologists who can diagnose brain tumors is not sufficient, because of the necessity of highly specialized skill. Computer-Aided Diagnosis by computational image analysis may dissolve the shortage of experts and support objective pathological examinations. Our purpose is to support a diagnosis from a microscopy image of brain cortex and to identify brain tumor by medical image processing. In this study, we analyze Astrocytes that is a type of glia cell of central nerve system. It is not easy for an expert to discriminate brain tumor correctly since the difference between astrocytes and low grade astrocytoma (tumors formed from Astrocyte) is very slight. In this study, we present a novel method to segment cell regions robustly using BING objectness estimation and to classify brain tumors using deep convolutional neural networks (CNNs) constructed by deep residual learning. BING is a fast object detection method and we use pretrained BING model to detect brain cells. After that, we apply a sequence of post-processing like Voronoi diagram, binarization, watershed transform to obtain fine segmentation. For classification using CNNs, a usual way of data argumentation is applied to brain cells database. Experimental results showed 98.5% accuracy of classification and 98.2% accuracy of segmentation.

A multilevel surface modeling method and its application to range image registration

This paper proposes a new method for representing 3D shapes based on spatial approximation by spheres with discretized radii. The shape representation is composed of several levels of spherical expressions that characterize the spatial resolution of a given model. The radius of spheres in each level of expression is determined based on a volumetric primitive which is defined by bounding the given shape. This multilevel surface modeling is useful not only for hierarchical model operations like collision detection, database retrieval and so on, but also for hierarchical partial shape matching with a portion of object surface. We applied our shape representation method to range image registration by an ICP algorithm. As a result, it was proved that the method is very effective for model matching and registration, and also its computational cost is remarkably reduced compared with the original ICP algorithm.

Pulmonary motion tracking from 4D-CT images using a 3D-KLT tracker

We propose a new method for lung-motion tracking and its quantification from 4-dimensional X-ray computed to-mographic (4D-CT) images. This method uses an enhanced 3D-KLT tracker. An advantage of our method is that it can find many feature points (regions) for tracking that are not restricted to the bifurcation points of bronchi or vessels. The feature point extraction algorithm depends only on image gradients. Moreover, our method adopts a hierarchical tracking based on pyramidal image structure. This provides robustness for large movements of the objects. Lung motion is quantified by tracking a large number of feature points in the lung. In this paper, we first evaluate the performance of our proposed method for artificial 4D-CT images and then describe quantification results of real 4D-CT images. Our experimental results clearly show that lung movement is not modeled by a simple translation but by an oval pattern.

Constraints Optimization for Minimizing Stretch in Bounded-Parameterization

In a mesh parameterization process that generates one-to-one mapping information between a three-dimensional surface and a two-dimensional plane, we need to set some constraint positions in the solving system to define a specific location or size of the mapping plane. In this study, we will discuss how to perform a bounded-parameterization that minimizes distortion based on changing of constraints setting in the solving system. We introduce a series of experiments focusing on constraints optimization to deliver the optimal mapping information with less computational cost and time. Our proposed methods can reduce more than half of calculation costs and times from traditional method while maintaining the optimal mapping information.

A Homotopy Surface Cutting using Paths Crossing in Geodesic Distance

Topology is a property of surfaces that plays a major role in computer graphics. Processing or analysis between two surfaces generally requires both of them to be in same topology. There are many tools or applications such as parameterization or remeshing that require disk topology surfaces as input. Therefore, it is important to convert any surfaces to be same as a topological disk. The common procedure is to define a graph of edges inside the surface that should be split into two edges and to turn the surface into topological disk. We call it as homotopy cutting. Problems become more difficult when dealing with high genus surfaces such as a torus. Based on a novel method, we present an enhancement method to generate a cut graph in high-genus surface for homotopy cutting. By using geodesic properties of each edge, we can generate equally or more suitable edge-graph than original method while keeping similar performance and stability as original one.

Dense motion analysis and segmentation of ultrasound images

We propose a dense motion analysis method for ultrasound images. A motion analysis is implemented by tracking a lot of lattice points. In this paper, two novel processings are introduced to perform the motion analysis. One is the tracking of lattice points based on an optical flow algorithm in a framework of multiple spring-models. The other is the detection of lattice points based on texture information with confidence value, and its result corrects the tracking errors. We evaluated our method using a sequence of artificial ultrasound images up to 5 minutes. The average and maximum errors of our proposed method have achieved the best performance in the conventional methods.

Liver Segmentation Based on Reaction-Diffusion Evolution and Chan-Vese Model in 4DCT

Localization is an important step in the radiation treatment planning. The use of 4DCT data can enhance the efficiency of the planning when a target region is deformed by respiratory motion. Conversely, image quality in soft tissue is low since it utilizes low energy to collect data in order to limit the accumulated dose in a patient. This paper presents a method of liver segmentation in 4DCT data including high image noise and metal artifact. The proposed method was based on a level-set method using reaction-diffusion evolution and modification of a Chan-Vese model. Automatic segmentation was independently performed on each CT volume in a breathing cycle. From our results, the global shape of the liver was extracted smoothly without detecting extraordinary regions. The displacement computed from the center of mass of the liver-segmented volume was similar to a movement trend of two metal markers placed inside the liver.