Shop

Showing 13–24 of 161 results

  • Magneto-optical images for nondestructive inspection of plant steel structures using deep generative models

    Abstract:  Measures against deterioration of infrastructures that were built during the high economic growth period are facing significant challenges with regard to the maintenance of infrastructures in Japan. The development of optimal nondestructive sensing and imaging technology according to the material and structure of buildings is underway to contribute to efficient and reliable maintenance of infrastructures. However, owing to the large number of materials and structures used for buildings, as well as the types of defects to be targeted, many basic studies are yet to reach the stage of practical use. In this study, we developed a magneto-optical (MO) sensor in order to visualize a “crack” in the plant steel structure and automatically detected the defects in the plant steel structure by performing deep learning on the MO image obtained. As a pretreatment for detecting anomalies in defects using the AI, we focused on the nondestructive inspection using MO imaging and performed an unprecedented image filter processing. As a result, automatically evaluation the several types of MO images using AI, the accuracy of defection identification was improved.

    Keywords: artificial intelligence; variational autoencoder, nondestructive inspection; magneto-optical imaging

  • Object Searching Robot Controlled by Edge-AI

    Abstract:  This study proposes and develops an edge-AI-based autonomous mobile robot based on open-source software. The robot is capable of voice and object recognition; it can detect and approach an object specified by a user’s voice. Because the robot is controlled by voice commands, the user can control the robot intuitively. In the present study, we used a robot operating system to facilitate the development. All functions, including voice recognition, object recognition, and motor control, were implemented in the edge AI computer based on open-source software. We conducted preliminary experiments to verify the performance of the developed system.

    Keywords: Mobile robot, Edge-AI, Open-source software, Image recognition

  • Proficiency Estimation Method of Vibrato in Electric Guitar

    Abstract:  Many systems that provide automatic evaluation and feedback of electric guitar have been developed. However, they have a serious weakness, only “timing” and “pitch” are considered in evaluation. In fact, a wide range of factors are involved in the human evaluation. In order to solve this problem, previous studies proposed automatic evaluation methods. On the other hand, these are not possible to evaluate the sound using the special technique of electric guitar. In this study, we proposed a method for automatic proficiency estimation of vibrato in electric guitar. As the method, we extracted the acoustic features focusing on peaks of Mel fundamental frequency, number of peaks, width average, width variance, height average and height variance. We regressed the evaluation values using extracted acoustic features with a relevance vector machine; RVM. As the result, we were able to perform regression with a coefficient of determination 0.723. This result indicates extracted features are highly relevant to evaluation values by human and allows a tentative evaluation of the vibrato sound of the electric guitar.

    Keywords: Electric Guitar, Vibrato, Proficiency Estimation, RVM, Mel Fundamental Frequency, Audio Signal Processing, Machine Learning, Music Information Processing, Music Education

  • Gastric Cancer Detection by Two-step Learning in Near-Infrared Hyperspectral Imaging

    Abstract:  Gastric cancer is one of the most serious cancers that affects and kills many people around the world every year. Early treatment of gastric cancer dramatically improves the survival rate. Endoscopy has become an important tool for early detection. Since invasive gastric cancer or the edge of the invasive gastric cancer is difficult to find by using conventional visible-light endoscopy, near-infrared imaging, which is bringing great progress to the medical field, is focused on in recent years.  In order to apply near-infrared hyperspectral imaging (NIR-HSI) in real-time, wavelength feature extraction is important because a large amount of data needs to be analyzed. The purpose of this study is to detect gastric cancer using NIR-HSI and to select a suitable wavelength for the target in the near-infrared region (1000–1600 nm). NIR-HSI was used to take data from six specimens of gastric cancer and each pixel was labeled as normal or tumor on the hyperspectral image based on the histopathological diagnosis. 4 wavelengths were extracted from 95 wavelengths using the least absolute shrinkage and selection operator method. Supervised learning was performed using a support vector machine for both cases using all 95 wavelengths and the case using 4 selected wavelengths. In both cases, the approximate location of the tumor could be identified, indicating that an appropriate wavelength could be selected. We were also able to improve the detection accuracy by creating new supervised data and adding another learner. The detection accuracy was 93.3% for accuracy, 69.8% for sensitivity, and 96.7% for specificity. These results show that gastric cancer can be detected even at four wavelengths. By applying the results of this study to the endoscope system, the possibility of constructing a NIR endoscope system for gastric cancer was suggested.

    Keywords: Cancer detection; Gastric cancer; Near-infrared hyperspectral imaging; The least absolute shrinkage and selection operator; Support vector machine

  • Non-Predefined Life Signs Detection for Disaster Survivors Rescue

    Abstract:  No one can tell or predict when and where a natural disaster such as an earthquake or tornado will occur and the damages they may cause. Or an overflow of a large amount of water beyond its normal limit (water flood) shallowing city as we used to watch on news TV after the passage of a strong typhoon causing heavy rain. These natural disasters occur every day and anywhere around the globe are not new. And we cannot not prevent them from occurring in spite of the best technology we have in our possession now. But saving lives after their occurring is still possible and the best technology for this is the combination of AUV and the image processing. Image processing is one of the best ever invented technology by human since the course on technology development between scientists for sustainable development of our society. In order word “image processing is the technology that meets the needs of the 21st society we live in without compromising the ability of future generation to meet what they need to make the use of this technology efficiency. This paper proposes non-predefined life signs detection for disaster survivors rescue when a disaster occur and especially during a floodwater. In this research we use the matrix-based pairs of opposing pixels positioned directly around the observed point that belongs to the edge of the life signs target.  At first one-dimension matrix for bitmap memorization values of the RGB components of pixel is used. Next these values of the RGB components of pixel color are copied from bitmap to matrix. The number of bytes in a row is rounded up to the nearest number divisible by four. As a result, the method clearly detects all life signs edge made by human using any type of item around them with 95%.

    Keywords: Life signs; Image processing; Disaster; Edge detection; Rescue

  • Seashore Debris Detection Model with KaKaXi Camera Custom Dataset Using Instance Segmentation

    Abstract: Marine debris is impacting coastal landscapes majorly by affecting biodiversity, impairing recreational uses, causing losses to fishing industries, maritime industries, etc. Motivated by the need for automatic and cost-effective approaches for debris monitoring and removal, we employed computer vision technique together with deep learning-based model to identify and classify marine debris on several beach locations. This paper provides a comparative analysis of state-of-the-art deep learning architectures and proposed architecture which is used as feature extractor for debris image classification.

    The model is being proposed to detect seven categories of marine debris using a custom debris dataset, with the help of instance segmentation and a shape matching network, which can then be cleaned timely and efficiently. The manually constructed dataset for this system is created by annotating fixed KaKaXi camera images using CVAT with seven types of labels. A pre-trained HOG shape feature extractor is being used on LIBSVM along with template matching to improve the predicted masked images obtained via Mask R-CNN training. This system intends to timely alert the cleanup organizations with the recorded live debris data. The proposed network resulted in the improvement of misclassification of debris masks for objects with different illuminations, shape, occlusion and viewpoints.

    Keywords: debris; fixed camera images; computer vision; instance segmentation; deep learning; template matching; Histogram of Gradients (HOG)

  • Quantitative Evaluation of Orthodontic Treatment by Moment Measurement Device

    Abstract: The purpose of this study is to quantitatively evaluate orthodontic treatment. The forces and moments applied to the teeth during treatment are rarely measured. Therefore, dentists must rely on their own skills, experience, and senses to perform treatment, which may not be sufficient for some patients. To solve this problem, devices have been developed to measure the forces and moments generated in the teeth. However, there are disadvantages, such as the limitation of the direction in which the forces and moments can be measured and the fact that they do not consider the movement of the teeth during treatment. Therefore, we have developed a device that can measure forces and moments in all axes, reproducing the movement of the teeth during treatment. The developed device consists of two teeth model, a force sensor, and a stepping motor. Considering that the teeth move during treatment, this device was incorporated a motor to control the angle of the teeth. A force sensor is used to measure the forces and moments in the three axes applied to the teeth. This device can reproduce the treatment of abnormally tilted teeth until teeth return to normal position. Quantitative evaluation was performed using the device. A comparative study was conducted between the case of treatment with stainless steel wires and the case of treatment with nickel titanium wires. The result showed that nickel titanium wire has many advantages compare the conventional stainless-steel wires. The advantages are agreed with the dentist’s rule of thumb, and the experimental results suggest that this device was able to make a quantitative evaluation of orthodontic treatment.

    Keywords: Force and moment measurement device; Movement of teeth; Orthodontic treatment

  • Emergent Transition from Radial Foraging to Tree-like Raiding Patterns Induced by Complete Following for Pheromone

    Abstract:  Among ants that show complex and diverse collective actions, army ants are known to raid prey in groups. It has been confirmed that the characteristic pattern in a swarm raid of army ants changes from radial to tree-like with time. There are some simulation studies focusing only on the emergence of tree-like patterns. However, the models adopted in their simulations cannot represent the transition of patterns from radial to tree-like observed in the real world. In this study, we propose an army ant model with a modification of the model adopted by Solé et al. and clarify the conditions of the emergence of the transition. From the experimental simulation results of our model, we deduce that the simple modification of the Solé model can move individuals in eight directions, but it is not enough to express multidirectional to unidirectional convergence. Additionally, our simulation experiments showed that for the pattern transition from radial to tree-like, it is necessary for the individuals to keep moving forward until they get food, and for the returning individual to completely follow the pheromone.

    Keywords: army ant model, radial foraging pattern, tree-like raiding pattern, swarm raid, complete following for pheromone.

  • Speech Recognition Signal Lamp Image Simulation

    Abstract:  Speech recognition technology is a method of computer sound signal processing, which determines human behavior by analyzing the characteristics of sound signal. It has a wide range of applications in modern science and technology, and is a new frontier science, also is known as intelligent language. With the continuous improvement of people’s requirements for the quality of life, intelligent devices have sprung up in various fields. It is very practical to apply speech recognition technology to smart home reasonably to make home life more comfortable, safe and effective. The recognition of speech signal is primarily completed by preprocessing, feature extraction, training and pattern matching; the user interface is established by using the function of Matlab GUI, and the signal lamp image based on speech recognition is simulated and controlled by using the software.

    Keywords: Speech recognition; Endpoint detection; Feature extraction; pattern recognition

  • Intelligent Ramp Patrol Car based on MSP430

    Abstract:  This article aims to use the MSP430 single-chip to design an intelligent car, which is suitable for the precision tracking of various inclined slopes. The patrol car uses infrared sensors to collect ramp trajectory information and to adjust the forward direction through the front-axle steering mechanism. The gyroscope collects the car status in real time so that the car is controlled to perform uphill acceleration, downhill deceleration. Such a design allows the patrol car to be stabilized in the ramp with the established route.

    Keywords: MPU6050, patrol car, front steering axle, ramp track

  • Research on Similar Odor Recognition Based on Bid Data Analysis

    Abstract: In the common olfactory system odor recognition is processed by the electronic nose collecting sensor data, but the odor data collection of substances is easily affected by the environment and the processing is complicated, which is prone to deviation. This paper proposes a method based on big data analysis. According to the different chemical structure characteristics of different odor substances, the BP neural network is used to build a model to classify and recognize similar odors, and compare it with the traditional PCA+LDA recognition method. The results show that the establishment of a similar odor recognition model can accurately classify substances with similar odors, and the BP neural network algorithm is used to identify different substances with a higher rate of odor recognition. This method is stable and simple, and can provide different ideas for odor identification.

    Keywords: Smell recognition; Big data analysis; BP neural network; Similar smell

  • Text Classification Based on Title Sematic Information

    Abstract: With the rapid development of big data technology, text classification plays an important role in practical application, its applications span a wide range of activities such as sentiment analysis, spam detection, etc. Traditionally, we model the relationship between document and label. However, in many scenarios, document have specific relationship with corresponding title. Inspired by this, a text classification model based on title Semantic Information is proposed in this study. In our model, long short-term memory (LSTM)is used to learn title embedding, document embedding is obtained by using promoted LSTM(TS-LSTM) which take into account the title information. The experimental results on the standard text classification datasets show that its performance is better than the existing state-of-the-art text classification algorithms.

    Keywords: Text classification; natural language processing; deep learning; LSTM