In this paper, as examples of applications of 3D image measurement, we introduce applications to medicine/health, agriculture, and media art and propose a food texture camera using high-precision 3D measurement as a new application. In addition, we will also show the prospects for cooperation and internationalization with the Institute of Technology of Cambodia, which has started working as one of the goals of international dissemination of 3D image measurement technology.
Nowadays, flexible array probes are actively studied for ultrasonic non-destructive inspection of various shapes. However, in order to perform imaging with ultrasound, it is necessary to know the shape of the probe at the time of imaging. Conventionally, a method such as attaching a surface shape measurement device to a flexible array probe has been used, but there are problems such as the cost and labor involved in using a separate device. Therefore, in this research, we developed a self-sensing flexible array probe that can measure its own shape using the piezoelectric effect of the array sensor. The principle of self-sensing is the piezoelectric effect of the ultrasonic element, and the voltage output when the element is deformed is measured by time integration. In addition, we devised a shape estimation method at an inflection point where the curvature changes in shape measurement, and tried to improve the shape measurement accuracy. As a result, the positional coordinates of the element could be estimated with an error of 0.5 mm from the true value, and the response speed was about 200 ms or less.
Standard insertion machines require pre-determined component position and posture. If they change every time, we must solve this problem. Most conventional methods attempted to solve this task by identifying the position and posture. However, these methods require a multi-step strategy following the handmade rule. This paper proposes an imitation learning method to automate the wire insertion task with uncertainties in position and posture. The proposed model learns the motion policy through human demonstrations and maps image data to the robot's action in a single step. Moreover, the model considers the parallax of the stereo images for accurate insertion. In addition, the model outputs the insertion action and recovery action to recover from insertion failures. However, the standard data collection method cannot collect recovery actions, and manual labeling of action classes is essential. This paper proposes a novel data collection method called "Labeling with Human Intervention (LHI)" to tackle this problem. This method automatically generates action labels and collects recovery action with human intervention. We conducted real-space insertion tests and found that our approach achieved 97.2% (35/36).
In this paper, we describe head detection and pose estimation of infants sleeping in a nursery school. Every year, there are many cases of infants dying from put in a prone sleeping position during nap time. Because a nursery teacher cares for several infants, and even if they pay attention, they cannot check on the condition of all the infants. We have been working on the development of a system that attaches a camera to the ceiling to take video of infants, detects infants of face down position, and sends alerts to the smartphone of a nursery teacher. Pose estimation focusing only on the infant's head cannot use existing methods for estimating face orientation because the infant's face may not be in the video. Therefore, we propose a two-step method for head detection and head orientation estimation. For head detection and head orientation estimation, YOLOv5 and VGG-16 with deep learning are employed to create a head detector and a head orientation estimator. Based on the experimental results, we will compare the accuracy of head detection and head orientation estimation with conventional methods, and verify their usefulness in the system.
In order to realize a life support robot that can be handled universally, it is necessary to generate robot behaviors that flexibly handle various tools according to diverse tasks. However, there are countless tools in our living environment, and it is difficult to teach each tool its own motion. We propose method to automatically generate robot motions for tools of unknown size and shape to achieve their intended use. I propose the available Common Motion Trajectory Model for the several kinds of tools which a purpose of use was common to commonly. Furthermore, I propose method to generate the robot motion that can execute a purpose task appropriately by reflecting the function information of the tool which sensing in Common Motion Trajectory Model. In the experiment, we performed the task of “scooping and pouring matcha” for a total of 15 different utensils in three categories (Spoon, Ladle, and Spatula), and confirmed that the average success rate was 71.0%.
Saliency maps for predicting areas of human fixations from images have been actively studied. In order to improve the accuracy of the video saliency model, we compared the saliency map predicted by video saliency models with the saliency map measured by experiment. We found that the area of fixations is biased in the direction where the target is moving and the target cannot be gazed at when the motion of the target changes. Moreover, we confirmed that the findings are unavailable for one of the state-of-the-art video saliency models, STSANet.
Parasitoid wasp is used for pest control due to by using enemy relationship. However, it is essential to identify them because parasitic host is different for each parasitoid wasp if they use parasitoid wasps for pest control. Currently, traits extraction, measurement, calculate and create specimen used for identification. However, there process is burden for parasitoid wasp researcher. In addition, they are shortage of parasitoid wasp researcher while the burden is heavy. Therefore, it assists parasitoid wasp researcher extracting node that is vein linked used in identification by using deep learning and image processing in this research.
The Capsule endoscopy is a technique to capture images of the inside of the gastrointestinal tract by swallowing a device measuring approximately 11 mm in diameter and 26 mm in length. Compared with conventional endoscopy, capsule endoscopy is less burdensome on patients while allowing observations of the small intestine. This non-invasive technique produces more than 50,000 images in a single examination. Because a physician must visually check each image, a diagnosis is time consuming and labor intensive.
This study investigated automatic detection of lesions to reduce the burden on physicians, preventing missed lesions and support diagnosis.
Here, we use YOLOv5 (You Only Look Once version 5), which is a general object detection model, to automatically detect lesions after training a model with 3 types of lesion images. When the recall was 100% to ensure that no lesion was missed, the polyp accuracy, ulcer accuracy, Type A accuracy and Type B accuracy were 96%, 99%, 77% and 94%, respectively. In the future, we will train the model with additional images of other lesions and improve the precision rate.
Amusement facilities such as pachinko parlors have been suffered from one of the social problems where a part of the customers got stressed out by failing their games and carried out destruction of the facility such as lavatories. Solving the problem by visual surveillance using monitoring cameras is not ideal from the viewpoint of privacy. Instead, detecting abnormal sounds which come from those destruction could be the alternative. This work collected all the sounds that could be generated in a public lavatory and analyzed their acoustic characteristics to obtain their statistical distribution. A proposed method of discriminating abnormal sounds utilizes the statistical distribution of the acoustic parameters using Gaussian mixture models to describe them mathematically. An experimental result regarding discriminating abnormal sounds demonstrated the significance of the proposed method.
This study deals with the task of segmentation of SEM images of fine ceramics sintered bodies by using Deep Neural Network (DNN). In particular, we focus on misclassification caused by the blurriness of grain boundaries(boundaries between particles). Therefore, we utilize the frequency distribution of brightness gradient of grain boundaries and give higher weights to pixels with lower gradient values. Experiments confirmed that the model trained with proposed loss function gave the best prediction results.
Respiratory rate (RR) is known to be a more accurate predictor of clinical deterioration than other vital signs. However, there were few respiration measurement devices certified as medical devices that could be used in daily clinical settings. Therefore, using a bult-in microcontroller (72×57×12 mm) instead of a personal computer, we developed a portable stand-alone respiration measurement device with minimum workload in computing that can be used for non-contact measurement in 30 seconds using a Doppler radar.
In this study, the problems of respiration measurement using a Doppler radar, such as miscounting of respiratory peaks were clarified, and proposed device with respiratory peaks miscount prevention algorithm achieved high accuracy RR measurement. Clinical testing was conducted on pediatric outpatients of Children's Medical Center, Showa University Koto Toyosu Hospital. The measurement accuracy of the system was confirmed to be comparable to the respiration measurement accuracy of stationary certified medical devices used in hospital, such as capnometers and the chest wall impedance method used in bedside monitors.
We propose a modeling framework for automating batch processes operation. Batch processes are often controlled by PID controllers, where engineers manually regulate their parameters and temporal patterns of reference signals. Therefore, it takes a long time for optimizing these parameters and temporal patterns. A possible solution for this is to apply so-called Model Predictive Control (MPC) technology to the tuning. Here, batch process dynamics depend on the types of products and of equipment, thereby forcing engineers to construct and maintain multiple models that correspond to the number of combinations of product types and equipment types. Thus, batch process modeling is a time-consuming and complicated task. To solve this problem, we propose a modeling framework; about a modeling target, the part applying commonly and parameters can be decided in advance are constructed by mathematical models, and the part that required experimentation for designing or tuning are constructed by machine learning models. We expect this framework can improving estimation accuracy and suppressing the number of model construction by separating model construction and combining the mathematical and machine learning models. In our simulation, we confirmed that our proposed model can suppress prediction error (RMSE) of reactor temperature under 1K. Furthermore, an optimization algorithm with our model can find a temporal pattern of a reference signal so as to reduce control error of reactor temperature under 1.99K.
This paper discusses the effect of observation noise in data-driven control. Generally, a set of input/output signals is used to calculate data-driven control (DDC). On the other hand, VIMT only uses the output signal without the input signal. Then, it seems that VIMT is affected by adverse effects of the observation noise compared with other DDC methods. In this paper, we consider the effect of observation noise against VIMT. Furthermore, we propose a simple solution to suppress the degradation of VIMT update performance due to observation noise. Finally, we verify the results using numerical and experimental examples.
Speech recognition systems have been applied in various fields due to recent development of digital signal processing techniques. For speech recognition in real circumstances, some countermeasure methods for surrounding noises are indispensable. In previous study, we derived an algorithm to estimate speech signals by using air-conducted sound mixed with noise as observation and using bone-conducted sound. At that time, a model of the observed air-conducted sound was represented by a simple additive model of the sound pressure level of the speech signal and the noise. In this paper, a signal processing method to estimate the speech signal is proposed by observing air-conducted speech under existence of surrounding noise and using bone-conducted speech. More specifically, an estimation algorithm reflecting higher order statistics of variables is derived by applying Bayes' theorem after introducing a stochastic model for the speech signal, air-and bone-conducted speeches. Furthermore, by applying the proposed method to speech signals measured under existence of surrounding noise, its effectiveness is confirmed.
When a disaster occurs and causes people to evacuate, delivering relief supplies to shelters is indispensable. In addition, the disaster may cause a large-scale and prolonged power outage; an adequate power supply to shelters is also necessary. In this paper, we suppose a situation where a disaster occurs in an area where a photovoltaic power generation system and storage batteries are installed in each shelter, and consider the delivery-planning problem of electricity and relief supplies by electric vehicles. First, we formulate the problem as a mixed-integer linear programming problem. Because the problem is an extension of the vehicle routing problem, which belongs to NP-hard, it is challenging to find an optimal solution for practical-scale cases. Therefore, we propose a heuristic method based on adaptive large neighborhood search and mathematical programming. Then, we conduct computational experiments to evaluate the performance of the proposed method. The results of the computational experiments show the effectiveness of the proposed method.
In the development of blink input interfaces, it is important to classify between conscious (voluntary) and naturally occurring (involuntary) blinks. In the previous studies, some systems employed a long blink as a voluntary blink, but determining the appropriate discriminative condition was difficult. To avoid the problem of individual differences in discrimination conditions, an individual calibration method was proposed, but the calibration procedure increases the burden on the user. In this study, we introduce a new 3D convolutional neural network (3D CNN), which deals with spatial and temporal dimensional directions. This 3D CNN model is trained with a moving image dataset of the periocular area. For the proposed 3D CNN, a sequence of seven images cut from a video sequence is used as a set of an input sample to classify three states; voluntary, involuntary, and not blinking. In this study, data of five subjects were used for training and seven for testing. A detailed analysis of the result revealed that the biased position of the open-eye area in the images leads to a lower classification rate. To address this problem, we propose an automatic determination method for the area to be cropped in the periocular image and verify its performance.
Soft tennis originated in Japan, is selected as a club activity by many junior and high school students. Several senior elite-level competitions, e. g., nationwide amateur club championships, have been held regularly. These competitions, however, have no clear order of importance between them. Moreover, any ranking of players or pairs based on mathematical evidence has not been established. This paper proposes a novel ranking method for elite-level soft tennis players using match results. The proposed method is based on Colley's method, which can include the opponent and the margin of victory of every match in strength evaluation. The validity of the proposed method is verified based on its prediction performance.
In previous research, we estimated snow cover rate on the panel from panel image for the purpose of improving the prediction accuracy of the PV power considering the effect of the snow accumulation on the panel in the snowfall area. A method to predict the quantity was proposed. However, in previous research, the snow cover rate was estimated by dividing it into 11 categories from 0% to 100% in increments of 10%. There is a question as to whether or not it is possible, and we think that estimating with even finer granularity will lead to further improvement in the prediction accuracy of PV power. We will compare and evaluate the prediction of PV power generation using the snow cover rate predicted from panel image with related research and confirm the effectiveness of the proposed method.
This letter proposes a measuring algorithm of yaw angle using a stereo camera. The proposed method estimates the change of the yaw angle by aligning the point groups with respect to the result of the three-dimensional reconstruction of consecutive frames obtained from the stereo camera. The algorithm prevents the degradation of 3D point cloud correspondence. Experimental results of head posture angle show the effectiveness of the proposed method.