-
Michito TSUCHIYAMA, Yuukou HORITA
Session ID: 22-03-01
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Nystagmus, which causes the eyeballs to move or shake spasmodically and independently of the willpower, may occur during everyday vision, or it may be caused by abnormalities in the brain, ears, or other parts of the body. Currently, videoculography (VOG) is used to diagnose nystagmus, but current analysis methods have problems with misrecognition of the pupil due to the narrowness of the eye slit, the pupil being obscured by long eyelashes, and the influence of blinking, which affect stable nystagmus detection. In this study, we developed an automatic nystagmus detection method using a combination of image processing and machine learning based on video images taken by magnifying the eye area, and evaluated its performance.
View full abstract
-
Rina TAKAGI, Yota YAMAMOTO, Ryosuke FURUTA, Yukinobu TANIGUCHI
Session ID: 22-03-02
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In the manufacturing industry, there is a shortage of human resources due to the declining birthrate and aging population, as well as a growing mobility of human resources. Video is an effective way to reduce the cost of human resource training and to efficiently communicate complex work procedures and skills that are difficult to verbalize. Video can communicate complex work procedures intuitively, but it is difficult to search for specific scenes or skip over unnecessary scenes. Since manual editing of work videos is time-consuming and labor intensive, there is a growing need to automate the analysis of work videos. The first step is to segment a video into action segments or to detect action boundaries based on the similarity of image features. In this paper, we propose an action boundary detection method using hand detection for unedited work videos taken by a fixed camera, such as cooking and parts assembly videos. By taking advantage of the fact that hand movements, such as picking up or moving objects, often occur at action boundaries, our method does not require a large amount of training data with action labels.
View full abstract
-
Ryoga TAKAHASHI, Yota YAMAMOTO, Ryosuke FURUTA, Yukinobu TANIGUCHI
Session ID: 22-03-03
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Quantification of inspection is an important issue in the inspection of finished products in the manufacturing industry. Even in highly automated factories, many inspections still rely on visual inspection. Door-closing inspections of automobiles are one of these inspections, which rely heavily on sensory inspections, and human error is inevitable. Although quantitative inspection methods have been developed, they are time-consuming and labor-intensive because the minimum energy required for door closing must be measured with a dedicated measuring instrument. In this study, we realize door-closing abnormality detection based on multi-modal deep learning to improve the efficiency of door-closing inspections. Based on the observation that there are differences in video and audio features between normal and abnormal doors, we utilize multi-modal information; that is, not only video but also audio features.
View full abstract
-
Tomonori SUGIYAMA, Akihiro NAGATA, Satoshi NEMOTO
Session ID: 22-03-04
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Anti-counterfeiting measures that use individual product differences to discriminate are effective against mass counterfeiting. However, there are problems in that unique ID assignment such as code information has low duplication resistance, and that special discrimination equipment is required to use existing artifact metrics. In this study, we focused on colored fibers used in anti-counterfeiting paper. By combining individual judgment based on fiber distribution and authenticity judgment of the paper itself by machine learning, it is possible to judge authenticity with a single smartphone. We aimed to achieve both simple judgment and duplication resistance.
View full abstract
-
Tomoya FUJII, Rie JINKI, Yuukou HORITA
Session ID: 22-03-05
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The social infrastructure, including roads and bridges built during Japan's period of rapid economic growth, is rapidly deteriorating, and there is a need to strategically maintain and renew the social infrastructure that is aging all at once. On the other hand, in road maintenance and management in rural areas, it is not realistic to increase the number of road management patrol cars or the number of specialized engineers engaged in road maintenance and management, and the reduction of management budgets and the shortage of engineers due to the declining birthrate and aging population are serious problems. In addition, in rural areas, it is difficult to conduct all road inspections by visual inspection, which is performed by expert road maintenance technicians, and an inexpensive, high-precision system that can automatically detect road surface damage through image analysis or other means is required. In this study, we construct a road surface damage detection model using YOLOv5, a machine learning algorithm capable of real-time.
View full abstract
-
Shunpei AOU, Yota YAMAMOTO, Kazuaki NAKAMURA, Yukinobu TANIGUCHI
Session ID: 22-03-06
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
As dairy farming is becoming larger and the workforce is becoming scarce, it is becoming more difficult to continuously monitor the behavior of dairy cows and accurately detect signs of disease and estrus. To solve this problem, surveillance cameras have been installed in barns to record the behavior of dairy cows. Tracking the behaviors of dairy cows in a barn, the amount of activity, and the frequency of drinking and eating will lead to early detection of disease and the sign of estrus. Conventional methods for dairy cow tracking detect dairy cows as rectangles with vertical and horizontal sides, which are not tight (including areas other than dairy cows), so their tracking accuracy is insufficient. This paper proposes a multi-camera dairy cow tracking method using rotated rectangles is proposed to improve the tracking accuracy.
View full abstract
-
Yuki FUKUDA, Munetoshi IWAKIRI, Takumi FUJIWARA
Session ID: 22-03-07
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Most current motion recognition technologies, such as motion capture, track feature points such as joints and analyze changes in their coordinates. However, such methods cannot be applied to creatures or objects with complex deformation patterns that are difficult to track feature points. In this report, we propose a method for analyzing deformation patterns using changes in the frequency distribution of shape feature values of a 3D point cloud in order to recognize motion of non-rigid objects for which feature point tracking is difficult. Experimental results show that the self-similarity matrix can be used to segment repeated deformation patterns, and that it can be applied to deformation pattern retrieval, period analysis, and process discrimination by creating a similarity matrix between the discovered pattern and the object to be analyzed.
View full abstract
-
Ayaka NANRI, Osamu SHIKU, Yuji TESHIMA
Session ID: 22-03-08
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
At present, humans need to sort fish species manually after landing at fish markets. Since the sorting process is a heavy workload, automation of the process is required. Although fish recognition methods using color images of fish and images of laser-irradiated fish have been proposed, there is currently no method for recognizing fish based on their entire three-dimensional shapes. In this study, we represent 3D shape of entire horse mackerels and mackerels as point clouds and investigate a recognition method using PointNet which is limited to rotation around the Z-axis only. First, we investigated the relations between the number of points used for recognition, accuracy and the recognition speed, and our findings showed that using 128 points resulted in the best recognition performance. Additionally, we compared our method to original PointNet and found that limiting rotation improved recognition. Moreover, we compared the recognition of point clouds, color images and depth images, and found that point clouds recognition had equal or higher than those of color images or depth images.
View full abstract
-
Mei KODAMA
Session ID: 22-03-09
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
With the rapid spread of 5G and changes in viewing methods, video viewing via network is becoming increasingly popular, and it is indispensable in daily life. In particular, Cache-typed video distribution methods such as, CDN, ICN/CCN, etc. have been studied for the purpose of network stability and shortening the acquisition time of video data for viewing. For example, the LFU method based on frequency is a weighting method for videos with a high number of accesses, and the efficiency decreases in a viewing model that does not allow re-viewing. In the LRU method, there is a problem that the efficiency decreases when the access distribution is switched. Therefore, in this study, the author focuses on improving the efficiency of video distribution and video management methods in fixed charge VoD services, which have periodicity of viewing as video viewing services. In particular, by utilizing the viewing history (past information) and the next viewing probability (future information) based on the behavior and preferences of the service user, a priority calculation method is proposed to improve the efficiency of the previous video management method of the cache server. By the simulation experiments on the transmission efficiency, and the cache efficiency is considered.
View full abstract
-
Masahiro ISHIKAWA, Naoki KOBAYASHI, Tokiya ABE, Akinori HASHIGUCHI
Session ID: 22-03-10
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Hyperspectral images are expected to be applied to pathological image analysis. In this study, we investigated the effect of infrared spectral images by converting HE-stained specimens to Grimelius-stained specimens that is positive for neuroendocrine granule, using U-net and evaluating the conversion accuracy. The results suggest that the estimation accuracy may be improved by adding not only visible light spectroscopic images but also infrared light spectroscopic images.
View full abstract
-
Ryo ONUKI, Junji YAMATO, Chanjin SEO, Jun OHYA
Session ID: 22-03-11
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
This paper, proposes a method for autonomous driving of a robot using deep reinforcement at a disaster site in a narrow space without a preliminary environmental map. RGB-D images are acquired from a camera attached to the robot and are inputted to a neural network of a deep reinforcement so as to determine the robot's action. Here, to deal with temporal relationship between images, a deep learning that can handle time-series data is also used. In addition, back function using the tether connected to the robot is exploited. As a result of experiments in simulation environment, success rate of the robot’s arrival to the goal is better than the conventional method. In addition, with the back function, a goal success rate of 98% was achieved.
View full abstract
-
Hiroki Urakawa, Kitahiro Kaneda, Keiichi Iwamura
Session ID: 22-03-12
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In recent years, the number of installed surveillance cameras has increased significantly, but most of them have not been effectively utilized. If video images are used to track people, they can be expected to be utilized in marketing and epidemiological studies. We have been conducting research on Artificial fiber Patterns, which is a technology to add information to textile patterns without causing visual discomfort. We aim to realize a privacy-conscious person tracking and flow analysis system by embedding Artificial fiber Patterns in clothing and reading the information captured by a surveillance camera. In this study, we verified the relationship between the coefficient for creating Artificial fiber Patterns and the accuracy when the color of the ink is changed, and achieved information embedding in a way that is more obfuscated than in previous studies, and that the embedded area is less likely to cause a sense of discomfort.
View full abstract
-
Shinya Abe, Kazuki Yamato, Satoshi Ito
Session ID: 22-03-13
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Compressive sensing has been applied to MRI to reduce imaging time, and the use of deep learning for image reconstruction has attracted attentions. Among them, adversarial generative networks (GANs) have been reported to have high reconstruction performance. In this study, the quality of reconstructed images is improved by introducing ensemble learning of Fresnel transform images to GAN. The Fresnel transform can create multiple image sets with different diffraction strengths. Therefore, ensemble learning can be achieved by training multiple networks with different Fresnel transform image sets. Reconstruction experiments showed that the reconstruction performance exceeded that of the conventional method.
View full abstract
-
Jian LIN, Taiga WATANABE, Yuki HASHIZUME, Keisuke HASEGAWA, Tsuyoshi M ...
Session ID: 22-03-14
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
A method has been proposed to convert noisy speech data into images and remove the noise using U-Net, one of the fully convolutional networks. The authors have already conducted experiments to remove various types of noise in addition to human speech using this method. Good results were obtained in all experiments. In this study, we assumed that a specific person's voice is emphasized during the meeting to record his/her voice. Alternatively, we thought that the voice of the emergency announcement speaker or the voice of the evacuation guide is emphasized to convert into the text to convey it to the hearingimpaired person. The authors prepared multiple datasets for training and created a speech enhancement model for a specific speaker's speech from multiple (up to 6) speakers. Then, it is confirmed that the enhancing speech of a specific person in mixed voice data can be possible by regenerating the voice.
View full abstract
-
Haruto OGAWA, Takashi IWASE, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 22-03-15
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
A contour tracking method suitable for 3D natural geometry has been proposed to support the creation of remains maps. In this method, a set of partial points corresponding to an edge region is extracted from an input point set, an arbitrary starting point is set, and a growth vector is obtained considering the tracking direction and variation with respect to the local region, and contour lines are sequentially tracked. In the previous study, the ratio of tracking direction to local variation was fixed at 1:1, but in this paper, this ratio is varied and the length of the tracking line segment corresponding to the local region is adjusted in two steps to attempt more flexible tracking. Experiments on actual archaeological data was conducted to verify how the tracking results change.
View full abstract
-
Takashi IWASE, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 22-03-16
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
A contour tracking method is proposed for edges of 3D point clouds representing natural shapes, with the aim of applying the method to the creation of posthumous maps. In this method, a sub-point cloud containing edge components is extracted from the input point cloud, a starting point is set, and contour lines are extracted sequentially by considering both directional components and variation with respect to the local region. In this paper, we discuss an extension of the least-squares method to the point set corresponding to the starting point of the line segment extracted from the partial point set in order to achieve smoother contour extraction, and verify what kind of tracking lines can be obtained by experiments on actual archaeological data
View full abstract
-
Shota SANO, Yuki HASHIZUME, Keisuke HASEGAWA, Yuusuke KAWAKITA, Tsuyos ...
Session ID: 22-03-17
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
There are many different sound sources in a room, and classifying these sounds has many applications, such as monitoring a living situation. The authors investigate a method for estimating each sound in the environment by converting timeseries data of indoor sounds into spectrogram images and using them as input for transfer learning to build a discriminative model. In this process, the amount of sound data that can be prepared in advance is limited due to the effort required for recording and the variety of data types required. As a result, there may be cases where sufficient classification accuracy cannot be achieved due to insufficient data for training. Therefore, this study proposes and applies a data augmentation method to improve classification accuracy when the number of data is limited, with the aim of classifying single and mixed sounds exist in a room, and describes the results of clarifying its effectiveness.
View full abstract
-
Mitsuyasu OKAMURA, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 22-03-18
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
When reconstructing a 3D point cloud from a video sequence, if a moving object exists in the frame, it remains in the reconstructed 3D model and the accuracy of the reconstruction of the entire point cloud is degraded. To solve this problem, we have studied a method of detecting moving object regions using optical flow and Detectron2, and masking such regions for 3D reconstruction. In this paper, we newly detect candidate regions of moving objects in the frame using the angle of optical flow, improve the accuracy of detecting moving object regions by using Detectron2, and examine the effect of masked images in 3D restoration. We also studied the effect of Euclidean clustering as a post-processing method for the resulting 3D point cloud.
View full abstract
-
Taishi FURUTA, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 22-03-19
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
When creating a three-dimensional point cloud from photographs, a method of creating a three-dimensional point cloud that includes cubes by placing them as indicators in advance has been developed. However, detecting cubes from a huge amount of point cloud data is computationally expensive. In this study, we propose a method to extract only cube regions quickly and accurately by filtering using DoN features calculated from the acquired point cloud and color information and demonstrate the effectiveness of this method.
View full abstract
-
Kei MORISHITA, Munetoshi IWAKIRI, Kiyoshi TANAKA
Session ID: 22-03-20
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In this study, we focus on reconstructing 3D models from video sequence using SfM-MVS (Structure from MotionMulti View Stereo), and investigate the method of selecting frames using pHash (perceptual Hash) as an image feature. Through simulation experiments, we compare the method of selecting frames using conventional optical flow with the accuracy and computational cost of the reconstructed 3D point cloud, and conduct a discussion.
View full abstract
-
Daiki YAMAUCHI, Makoto J. HIRAYAMA
Session ID: 22-03-21
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In recent years, drones have been widely used for surveying terrain and creating 3D models. High accuracy is required for surveying and 3D modeling. One of the advantages of using drones is that the angle and distance of view can be easily changed. In this study, we will conduct an experiment to see how the 3D model created by photogrammetry differs depending on the position of the drone, the number of photos taken, and the resolution of the photos. The video images obtained by using DJI's dji mini2 were used for photogrammetry. For photogrammetry, Agisoft's Metashape photogrammetry software was used. Comparisons of the 3D models showed that the highest accuracy was obtained at low altitudes, when the number of photos was large, and when the resolution was high. Future plans include programmed autopilot and LiDAR experiments.
View full abstract
-
Daisuke MIYAZAKI
Session ID: 22-03-22
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
This paper introduces computer vision researches using polarization. This paper explains the diffuse reflection and specular reflection, as well as the method to decompose those two reflection components. Polarization also can improve the visibility of bad weather when fog disturbs the sight. Surface normal of the object can be estimated from the orientation of the reflection plane detected from polarization analysis of specular reflection using the knowledge that the surface normal is included in the reflection plane. The surface normal of transparent objects is estimated using the raytracing method extended with Mueller calculus.
View full abstract
-
Yoshiki Funahashi, Haoran Xie, Kazunori Miyata
Session ID: 22-03-23
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Computer-generated images are highly dependent on texture in terms of realism. In this study, we focus on rust textures commonly found in the real world and propose a rust texture generation system using user sketches as input. First, the rust texture is converted into a control map to extract large-scale variations and nonlocal features. Then, the control map is converted to sketch data to extract global shape information. With the control map and sketch from the rust texture. we constructed the training dataset for texture image generation. A two-stage generative model is proposed to generate a rust texture with a complex structure from a user sketch. Using the proposed method, we implemented a prototype system that generates rust textures from user sketches, and evaluated the fidelity of the generated rust textures to the input sketches, their realism and diversity as rust textures.
View full abstract
-
Xusheng Du, Haoran Xie
Session ID: 22-03-24
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The generation of 3D shapes based on sketches is a challenging task because sketches provide sparse andblurredinformation. In this paper, we decouple the generation of 3D models into a composite part retrieval moduleusingasketch-based feature matching method, a sketch-based part generation module using a deep learning approach withimplicitfield representation, and an assembly module for part models to accomplish the task of generating 3D shapes fromusers' inputsketches. We also provide a module that allows users to optimize the assembled model manually. In addition, for noviceusers,providing a reference also enables them to draw a sketch that better fits the contour of the object. We also provideshadowguidance as background references in real time. By comparing our approach with previous works of MeshSDFandSketch2Mesh, it demonstrates from an objective perspective that our approach outperforms in providing details of themodel.Moreover, we evaluate the user experience and usability of our system and the performance of the providedinteractiveinterface from a subjective perspective by inviting participants to conduct a user study.
View full abstract
-
Yuki TSUCHIHASHI, Kohei TOKOI
Session ID: 22-03-25
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In Cel-Shaded Rendering, it is difficult to reflect the animator's intention because of the conventional method of shading by 3DCG. Therefore, we propose a rendering method that reproduces the animator's simplification of the shadows. The proposed method uses a shadow catcher that simplifies the original model in order to reduce unnatural shadows caused by the unevenness of the model. This aims to reflect the animator's intention when expressing the falling shadows.
View full abstract
-
Naoki Matsutani, Mutsuo Sano, Hiraku Matsuda, Sho Oi
Session ID: 22-03-26
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
Currently, cut vegetables are visually inspected for foreign matter, and a highly reliable automatic foreign matter inspection method is required to improve productivity and quality. However, vegetables contain pattern variations such as leaf veins and coloration, and it is not easy to learn normal products compared to standard industrial parts inspection methods that have small variations in normal products. In this study, considering the properties of foreign matter, we proposed a foreign matter image inspection method using a generative adversarial network that can learn the variation of normal vegetables, and verified its effectiveness.
View full abstract
-
Shuhei KODAMA, Reika YAGI, Tokiichiro TAKAHASHI
Session ID: 22-03-27
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
In cartoons and animations, there is a method of representing same shaped and highly dense objects by mixing the flatly painted part and the faithfully painted part of each object. The method to reproduce its abstraction using a 3D model has been proposed, but it has a problem that the degree of abstraction is difficult to control because the parts of the abstraction are selected manually. In this research, we tried to control the degree of abstraction by automatically determining the part of abstraction from parameters for the degree of abstraction.
View full abstract
-
Sota Mihara, Kazuhisa Yanaka
Session ID: 22-03-28
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
We discovered a new illusion in which line segments on a single-color background are perceived as if they are continuously extended when alternately displayed with their tonally inverted images for about 120 ms each. For example, image A has a white background and black line segments, and image B is the tonally inverted version of image A. Image A and image B are displayed alternately. Although a single line segment is acceptable, multiple line segments are placed in parallel to each other to increase the illusion. The background color can be red, blue, green, etc. instead of white to create the illusion. In addition, there are cases in which the line segments appear to contract rather than to extend. The mechanism of this illusion is unknown at present.
View full abstract
-
Akiko SATO, Haoran XIE, Kazunori MIYATA
Session ID: 22-03-29
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
The aim of this study is to enable researchers to produce attractive presentation materials and to show that such presentation materials may depend on the quantity of images. Therefore, we asked the conference attendees about the fastforward that they would like to go hear for a poster presentation and posters they thought were good. The results of the survey are analyzed in relation to the quantity of images in the presentation materials. As a result, the images of top selected poster group had more than 1.5 times amount than the other poster groups. Regarding the area of the images, the average value for the top selected poster group was that the image occupied more than 25% of the total poster areas, while the average value for the image area for the unselected poster group was 10%. These investigation results suggest that the quantity of images is important for creating good presentation materials.
View full abstract
-
Hiroshi UNNO, Kazutake UEHIRA
Session ID: 22-03-30
Published: 2023
Released on J-STAGE: January 31, 2024
CONFERENCE PROCEEDINGS
RESTRICTED ACCESS
This paper describes a display technique which enables us to invisibly attach signal of information to a displayed image on a flat panel display (FPD) by using color-difference modulation and to extract the signal of information from the displayed image captured with a video camera. Two experiments were carried out in this study. First, we assessed invisibility of the attached signal of information to the displayed image for the human eye. Then, we examined readability of the attached signal of that for video camera.
View full abstract
