Kazuki Saruta

3D Face Recognition Based on Local Curvature Feature Matching

In this paper, an approach based on local curvature feature matching for 3D face recognition is proposed. K-L transformation is employed to adjust coordinate system and coarsely align 3D point cloud. Based on B-splines approximation, 3D facial surface reconstruction is implemented. Through analyzing curvature features of the fitted surface, local rigid facial patches are extracted. According to the extracted local patches, feature vectors are constructed to execute final recognition. Experimental results demonstrate high performance of the presented method and also show that the method is fairly effective for 3D face recognition.

Deep Convolutional Neural Networks for All-Day Pedestrian Detection

Pedestrian detection is a special topic in computer vision and plays a key role in intelligent vehicles and unmanned drive. Although recent pedestrian detect methods such as RPN_BF [1] have shown good performance from visible spectrum images at daytime, they have limited study for near-infrared image at nighttime. Unfortunately, when the traffic accident happened at night, the pedestrian is one of the most serious victims. Recently deep convolutional neural networks such as R-CNN/Faster R-CNN [2, 3] have shown excellent performance for object detection. In this paper, we investigate issues involving Faster R-CNN for construction of end-to-end all-day pedestrian detection system. We propose an effective baseline for pedestrian detection both on visible spectrum images and infrared images, using a same pre-train Faster R-CNN model. We comprehensively evaluate this method, the experiment results presenting competitive accuracy and acceptable running time.

An end-to-end cells detection approach for colon cancer histology images

The qualitative and quantitative analysis of different types of histopathology images of cancerous tissue can not only help us in better understanding of tumor but also explore various options for cancer treatment. However, it is still a challenging task due to cellular heterogeneity. Deep learning approaches have been shown to produce encouraging results on image detection in various tasks. In this paper, we investigate issues involving Faster R-CNN for construction of end-to-end colorectal adenocarcinoma images analysis system. We experimented with different types of network for extract features, and analyzed the impact of time and accuracy. In addition, we optimize the various stages of the network training process. We have evaluated them on a large dataset of colorectal adenocarcinoma images, consisting of more than 20,000 annotated cells belonging to four different classes. Our results presenting competitive accuracy and acceptable running time. Prospectively, the proposed methods could offer benefit to pathology practice in terms of quantitative analysis of tissue constituents in whole-slide images. Code and dataset will be made publicly available.

An Image Inpainting Method Using Information of Damage Region

With the development of digital images processing, image inpainting is become one of the most impressive and useful technique. Based on partial derivative equations or texture synthesis, many image inpainting techniques have been proposed. Amano et al. proposed a method that based on eigenspace analyzes, which is called Back Projection for Lost Pixels (BPLP). It obtains the eigenspace from a set of learning samples from original image, then estimating the missing region by inverse projection and a linear combination of the eigenspace. But it remains have some obvious discomfort surrounding the damage region after restoration treatment. In this paper, we have proposed an adaptation of BPLP algorithms, which are performing well on images with natural defects.

Snow cover detection based on two-dimensional scatter plots from MODIS imagery data

Abstract. Snow cover detection (SCD) using remote sensing imagery has received increasing attention since the development of satellite remote sensing technology. In the present work, a SCD method based on two-dimensional (2-D) scatter plots generated from MODIS imagery data over Akita Prefecture in Japan is proposed. The imagery of the study area is preprocessed, including a geographic correction, clipping, an atmospheric correction, and a topographic correction, before SCD is conducted. For this, snow and cloud pixels are extracted from other ground surface features according to a 2-D scatter plot of bands 1 and 3 in the reflectance spectrum. Finally, a snow cover map of Akita Prefecture is obtained after removal of the cloud pixels detected from a 2-D scatter plot of bands 6 and 7. Comparison and validation with AMeDAS in situ snow depth data from the study area shows that the average accuracy obtained from our proposed method represents an improvement of 11.79% over the MOD10A1 product, and 22.05% over the SCD results from a combination of normalized difference snow index and normalized difference vegetation index. In addition, Aomori Prefecture and Mt. Chokaizan are also evaluated as further tests of the proposed method. All results suggest that the proposed method is feasible for SCD in the study areas and can provide information for agricultural development, water resource management, and ecological environment construction.

A Simple Visual Words Selection Strategy for Pedestrian Detection

An effective and efficient visual word selection method based on Bag-of-Features(BoF),which can be applied to the pedestrian detection problem in a single image, is proposed in this paper. We first calculate the difference in the total appearance frequency of each visual word in pedestrian and non-pedestrian images. Visual words that exhibit greater absolute values are more efficient for pedestrian detection, and are thus selected. The effectiveness of the proposed method is validated by analyzing the distribution of selected feature points. Through this analysis, we find that discriminative feature points for pedestrian images are mainly located about the lower body, whereas those for non-pedestrian images are mainly located in background areas. In addition, the experiments show that the time required for detection can be reduced by approximately 50%, with negligible loss in detection accuracy, using the proposed method, even if only 40% of the visual words are selected.

Comparison of FLAASH and 6S Code Atmospheric Correction on Snow Cover Detection in Akita Prefecture, Japan Using MODIS Imagery Data

In order to reduce the error, restore the real surface reflectance, in this work, FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) model and 6S (Second Simulation of a Satellite Signal in the Solar Spectrum) code model were used to conduct the atmospheric correction for Terra/MODIS data of Akita prefecture in northeast of Japan. From the results of snow cover detection, the accuracy of without atmospheric correction is only 46%, with FLAASH model for atmospheric correction is 86% and the accuracy of with 6S code model for atmospheric correction is 92%. As experimental results show, compared to the snow cover detection without atmospheric correction, the snow cover detection accuracy was improved by 40% and 46% with FLAASH model and 6S code model for atmospheric correction respectively. In addition, the accuracy with 6S code model is higher about 6% than FLAASH model.

Category classification with ROIs using object detector

Visual category recognition is challenging in computer vision and has several problem. Some of problems on visual category recognition are variance to the object instance position and background clutter. In this paper, we propose method select region of interest(ROI) in training and recognizing automatically. This provide invariance to object instance position and removing background clutter. In training phase, we make object detector to select ROI in recognizing automatically. The object detector is made by training regions of object and non-object, which determine a ROI without user annotation by using class label and some same class image of set of training image set. In this paper, the set of experiments is on the image database. We prove our proposed method can achieve high accuracy and recognize object position in training and recognizing

Computation Time of Grid Computing with Data Transfer Times that Follow a Gamma Distribution

The computation time of grid computing has rarely been evaluated for the case in which the computers used for calculation are scattered over the Internet. In the present paper, we assume that the data transfer times follow a gamma distribution and clarify the characteristics of the probability distribution of the computation time.

3D display system using monocular multiview displays

A 3D head mounted display (HMD) system is useful for constructing a virtual space. The authors have researched the virtual-reality systems connected with computer networks for real-time remote control and developed a low-priced real-time 3D display for building these systems. We developed a 3D HMD system using monocular multi-view displays. The 3D displaying technique of this monocular multi-view display is based on the concept of the super multi-view proposed by Kajiki at TAO (Telecommunications Advancement Organization of Japan) in 1996. Our 3D HMD has two monocular multi-view displays (used as a visual display unit) in order to display a picture to the left eye and the right eye. The left and right images are a pair of stereoscopic images for the left and right eyes, then stereoscopic 3D images are observed.