Effective radio propagation models are crucial for designing wireless communication systems that are both efficient and reliable, especially in tasks such as estimating coverage and analyzing interference. Mobile communication systems operate across diverse environments, including urban, suburban, and rural areas, necessitating propagation models tailored to specific geographical characteristics. We previously proposed a map-based method for clustering geographical data to classify environments. However, building a separate model for each cluster often suffers from limited training data, which can undermine the benefits of geographical classification. To address this trade-off, this paper introduces a novel approach: first, a global model is pre-trained using extensive training data that cover diverse geographical features across the entire area. Then, this global model is fine-tuned for each cluster identified through geographical clustering. This strategy customizes the propagation model to each geography while leveraging knowledge from the entire dataset, mitigating data scarcity in individual clusters. We validate the efficacy of the proposed approach using empirical data collected from various environmental settings. The proposed method demonstrates its effectiveness in reducing RMSE compared to a conventional method without fine-tuning, particularly in clusters representing high-rise areas, which are challenging to model due to propagation environment factors and limited training data.

For the advanced integration of cyberspace and physical spaces in Beyond 5G (B5G) and 6G systems, technical studies are progressing to achieve an ultrahigh capacity beyond what is offered by 5G, with a focus on terahertz bands. Although the 300-GHz band is particularly promising, the utilization of the terahertz band poses significant path loss challenges, necessitating an examination of its feasibility for use in mobile access networks. On the other hand, reports on radio propagation characteristics at the 300-GHz band are scarce, and even fewer focus on addressing nonline-of-sight (NLOS) conditions. Therefore, in this paper, we provide the results of propagation experiments conducted at the 300-GHz band in both indoor and outdoor environments, including NLOS conditions. We report the obtained path loss characteristics and angle of arrival characteristics. Furthermore, on the basis of the arrival wave distribution obtained from the angle of arrival measurements, we analyze and describe the propagation mechanisms of the 300-GHz band in NLOS regions.

This paper proposes a fast beam training algorithm utilizing true-time-delay (TTD) arrays suitably for high-rank line-of-sight (LoS) multiple input multiple output (MIMO) systems. In millimter wave bands, analog beamforming is generally utilized to compensate the effect of severe pathloss, and a beam training (BT) is typically operated to select beams which can obtain higher received signal power. However, it is difficult to secure the higher rank transmission in LoS-MIMO link with multiple sub-array panels, due to spatial correlation depending on a distance among Tx and Rx when selecting beam direction along LoS path commonly in all the panels. To alleviate this issue, we have proposed beamforming scheme, that select beams directing reflective paths at some panels, whereas the remaining panels select the direction of the direct path. However, it takes longer training time since BT is required to be operated in all the panels to select independent directions. Our proposed solution in this paper involves leveraging TTD arrays that can form a raibow beam, a set of multiple beams for each frequency, in one OFDM symbol, leading to reduction of the training time. Furthermore, we conduct effective angle-of-departure (AoD) and angle-of-arrival (AoA) estimation using TTD arrays to enhance the robustness of channel capacity for distance-dependent LoS-MIMO systems. The effectiveness of the proposed method is demonstrated through computer simulations. They show that the overhead of the proposed BT algorithm is smaller than that of BT algorithm using phased array (PA) while maintaining channel capacity.

Recently, methods have been explored to construct radio wave propagation models using machine learning, enhancing generalizability via transfer learning and fine-tuning. The conventional applications of machine learning are intended to estimate propagation environments within the learned area. Therefore, estimating propagation environments in unknown areas that are not represented in the training data is not expected and is challenging. This is presumably because, with limited data, the machine learning model does not adequately learn the physical propagation mechanisms. In this paper, we propose a path loss estimation method that classifies the dominant paths at each receiver based on physical propagation mechanisms. Specifically, we categorize the propagation paths into simple direct waves and single-reflection paths, as well as complex non-line-of-sight (NLOS) and estimate each accordingly. For line-of-sight (LOS) path loss, we use a statistical model fitted to measurement data to achieve easy and high-precision estimation. Additionally, for NLOS, we consider reflection loss based on a practical environment for single-reflection paths and apply a radar equation method that uses first Fresnel zone to estimate path loss accurately from close to distant ranges. For other NLOS, we propose a machine learning model that utilizes information from first Fresnel zone centered around the primary propagation path, estimating path loss at specific locations. The evaluation of the root mean squared error (RMSE) between actual measurements and estimates indicates that the proposed method improves the maximum RMSE by up to 11.5 dB compared with traditional machine learning approaches.

This research introduces a simplified beam training algorithm designed for line-of-sight (LoS) multi-input and multi-output (MIMO) systems, incorporating subarray-based beamforming capabilities at both the transmitter (Tx) and receiver (Rx), Typically, a conventional beam training algorithm is employed to maximize the signal power of each pair of Tx and Rx subarrays, leading to a scenario where each pair of subarrays directs its beam towards each other. However, in LoS-MIMO scenarios, the selected beams may not always be sufficient to achieve higher channel capacity due to spatial correlation dependency on the distance between Tx and Rx. To address this limitation, we have proposed leveraging reflection paths, such as ground reflection paths, and presented an effective beam training algorithm to implement this concept in real scenarios without requiring knowledge of the distance. In this paper, we propose a simplied beam training algorithm where the reflective direction determined by one subarray is shared with the other subarray. Experimental evaluation demonstrates for the first time that higher spatial multiplexing gain can be achieved without high dependency on the distance between TX and RX by controlling analog beam of each subarray appropriately under the existence of reflective paths.

In this paper, we propose a novel method to address “stamp artifacts”, a phenomenon in which unexpected objects are generated in diffusion model-based inpainting methods. A stamp artifact occurs when the region of an object to be removed is filled with a different object, resulting in unintended consequences. To solve this problem, we propose guiding diffusion model-based inpainting by incorporating results from traditional inpainting methods that use local information in the image. Experimental results showed that the proposed method effectively suppressed the appearance of unintended objects and achieved natural image completion.

With the development of real-time video streaming technology on the internet, live streaming services have been popular. In live streaming, it is common for users to have multiple live streams they want to watch simultaneously. Live streaming allows for simultaneous viewing across multiple windows. However, when engaging in simultaneous multi-window viewing, there is a risk of missing highlights from one stream while focusing on another. In this study, to avoid missing highlights during simultaneous viewing, we propose a method that predicts upcoming highlights by utilizing weighted moving averages that can capture trend changes in time series data in real time, based on chat text and audio from live streams. We evaluated the effectiveness of the proposed method using 23 YouTube Live streams, confirming that combining both textual and audio information enables effective real-time highlight prediction.

The Intelligent Reflecting Surface (IRS) is a promising technology for overcoming the coverage hole problem in millimeter wave communications; This equipment consists of many tunable passive elements and can reflect the incident wave in the intended direction. Physical optics (PO) approximation is used to calculate the path loss of the transmitter-IRS-receiver link. However, a higher computational cost is required as the number of passive elements increases. To address this issue, we propose a path loss estimation method based on the radar equation. Specifically, we have improved the radar equation using the first Fresnel zone within the IRS. Our proposed method can calculate the path loss more simply than PO approximation, less amount of calculation. In this paper, we show the effectiveness of the proposed method for arbitrary reflection positions with measurements and numerical simulations.

The realization of Cyber-Physical Systems (CPS) signals the potential resolution of various societal challenges. The exchange of diverse information between the physical and virtual realms necessitates the use of wireless systems, including those beyond 5G. In particular, the application of wireless emulators for the verification of wireless systems in virtual spaces is anticipated, enabling the efficient validation of systems with numerous wireless devices by replicating real-world communication environments in virtual spaces. However, accurate wireless communication emulation requires site-specific and high-precision radio wave propagation models. Recently, machine learning-based approaches have been proposed for modeling site-specific radio wave propagation characteristics. However, conventional models typically use training and test data from the same region, which can lead to degraded estimation accuracy when applied to unknown regions. This paper proposes a high-generalization model that improves estimation accuracy in unknown regions by using propagation path information as input features, based on the physical propagation mechanisms. The proposed model estimates a propagation path using machine learning and further uses this estimated path as input for machine learning to estimate propagation loss. When evaluating the estimation accuracy of propagation loss in unknown regions different from the training region, the proposed model reduced the RMSE from 5.52 dB to 4.69 dB compared to the conventional model, achieving an estimated error improvement of approximately 15%.

In Beyond 5G and 6G, the use of frequencies above 100 GHz is being considered to achieve ultra-high-speed and high-capacity communication. One application is to use these frequencies to connect peripheral devices such as wearable devices and user equipment (UE) to improve the communication speed of the uplink by using virtualized terminal technology. To evaluate the wireless performance of this technology, it is necessary to clarify the radio wave propagation characteristics and construct a propagation channel model especially for the Body Area Network (BAN). Therefore, we have studied estimation methods for diffraction loss in the 300 GHz band using a human phantom in an upright posture. However, in order to apply the estimation method to actual environments, it is necessary to consider various variables such as the effects of human posture and clothing. In this paper, we report the measurement results of propagation loss in the 300 GHz band using a real human body in a walking posture with clothes on, with a transmission point (Tx) on the right hand and a receiver point (Rx) at a comprehensive position near the human body, as an environment for actual use. In addition, we report that the propagation loss in the Line-of-Sight (LoS) region can be estimated with an accuracy of Root Mean Square Error (RMSE) 2.64dB using the Friis transmitting formula. Furthermore, we also report the propagation loss in the Non-Line-of-Sight (NLoS) region can be estimated with an accuracy of 8.98 dB in RMSE by combining the power of the diffracted path and the reflected scattering path.

In this paper, we report on scene classification experiments in FPS (First-Person Shooter) game videos focusing solely on acoustic information. Generally, scene classification in videos considers both visual and acoustic information. However, the visual information in FPS games contains game-specific elements that make it challenging to construct scene classifiers with generalization capabilities. In contrast, acoustic information in FPS games tends to have more commonalities, such as weapon usage sounds, which can contribute to successful scene classification. Focusing on acoustic information that provides essential clues for scene identification in FPS games, in the experiments, we employed a simple CNN model to classify audio segments into combat and non-combat scenes. To evaluate the classification performance, we used an FPS game audio test set. The test set consists of audio segments extracted from the audio streams of FPS game videos. Each audio segment is assigned a scene label indicating whether it belongs to a combat or noncombat scene. The results show that game scenes can be roughly classified with a simple CNN model by focusing on just acoustic information of FPS games.

Recently, deep learning-based image matting methods have emerged. However, the existing methods lack the capability to provide precise matting for anime-style illustrations because their network parameters are trained on primarily photo-realistic images. In this paper, we introduces a new anime image dataset, Chara-1M, designed for matting purposes. In addition, we propose AniCropify, a new matting method for character anime images. Focusing on the commonalities of representation between anime images and photo-realistic images, in AniCropify, an anime image is first converted into a photo-realistic image. From the converted image, a trimap is generated to identify the human regions in images. By using the trimap in the matting process, precise alpha masks of anime images can be obtained. From experiments, we confirmed that based on the quality evaluation of matting results, the proposed method received the highest rating compared to other state-of-the-art techniques.

This paper proposes a feature extraction method using the Hough transform as a method for utilizing map images in radio propagation estimation. The regression results of propagation loss parameters are evaluated in comparison with those by conventional methods using CNNs. Furthermore, we discuss how to extract feature parameters effectively.

In this paper, we propose a method for predicting radio wave propagation using a correlation graph convolutional neural network (C-Graph CNN). We examine what kind of parameters are suitable to be used as system parameters in C-Graph CNN. Performance of the proposed method is evaluated by the path loss estimation accuracy and the computational cost through simulation.

Advanced cyberspace and physical space integration is being pursued with B5G/6G, and technology studies for ultrahigh capacity beyond 5G are underway. The terahertz band, which enables ultrawideband transmission, is thus attracting much attention. However, this band has large path losses, so it is necessary to consider the feasibility of using it as an access line for mobile communication systems, and it is essential to understand its radiowave propagation characteristics. Among the terahertz bands with frequency over 100 GHz, the 300 GHz band (252-296 GHz), which ITU has agreed to be used for land mobile services with a bandwidth of 44 GHz, is expected to be utilized. However, there are few reports on propagation characteristics in this frequency band and few applicable radiowave propagation models. In this paper, we report the path loss characteristics obtained from radiowave propagation experiments at 300 GHz in indoor and outdoor environments. Additionally, we report the propagation mechanism of the 300 GHz band as revealed by arrival wave distributions obtained by angle of arrival measurements. Our findings indicate that diffraction contributes minimally to both indoor and outdoor environments, while reflected and scattered waves dominate the propagation characteristics. We also showed that use of a site-specific model, such as the P.1411 model that considers curved corners in outdoor environments, is effective for modeling the steep changes in the transition region in indoor and outdoor environments.

Positioning methods with an Intelligent Reflecting Surface (IRS) to enable control of the desired reflection direction has been studied. In the conventional Time of Arrival (TOA)-based positioning, the User Equipment (UE) is assumed to be in a coverage hole. In this case, the direct path between the Base Station (BS) and UE is completely blocked, and only the cascaded BS-IRS and IRS-UE paths can be obtained. However, since the UE could be in LOS of BS, there is concern that the error would be larger in this case due to differences in the characteristics of the paths. Therefore, in TOA-based positioning with IRS, it is necessary to identify the received signal as a direct or cascaded path. In this paper, we propose a TOA-based positioning scheme with IRS, including direct path/cascaded path identification to improve the positioning accuracy. Numerical simulations show that the proposed positioning improves the positioning accuracy compared to positioning approaches that do not consider whether the paths between the BS and UE are direct or cascaded.

Understanding the propagation characteristics in the real communication environment is crucial in designing and validating wireless systems. In particular, the path loss is one of the most fundamental and critical characteristics. Recently, machine learning technique has been used for site-specific path loss modeling. Several methods use building or satellite images as input data and apply the convolutional neural network (CNN) to model path loss according to the environment. CNN requires input data with fixed size due to its architecture. Therefore, these methods extract and input environmental information from a certain area around the transmitter and receiver points. As a result, accuracy degradation may occur due to redundancy or lack of environmental information depending on the distance between the transmitter and receiver. Therefore, this paper proposes a method for modeling path loss using a recurrent neural network (RNN), treating environmental information as sequential data. This method makes it possible to input the appropriate range of environmental information for various distances without redundancy or lack. Finally, we demonstrated the proposed method’s effectiveness by evaluating measurement data in the 2.1 GHz band acquired in an urban area. Especially in the short range, the evaluation results show that the proposed method can improve the accuracy by up to approximately 3.1 dB compared to the CNN-based method.

Path loss in the actual communication environment is one of the most fundamental characteristics in designing and validating wireless systems. Recently, machine learning-based path loss modeling methods have been proposed. Convolutional neural network (CNN) approaches have achieved site-specific modeling by extracting features of urban structures from building data, satellite images, and other sources. However, the previous studies have mainly focused on the frequency band below 6 GHz (Sub-6), and a few studies have applied this technology to the millimeter wave (mmWave) band. Due to its sensitivity to smaller structures, mmWave may require more detailed map data than the Sub-6 band. On the other hand, when a high-resolution map is input into a CNN, extracting the features of the entire urban structure could be challenging despite the representation of fine object information. This paper proposes a mmWave path loss modeling method using multi-resolution map data. This method improves modeling accuracy by incorporating the macro urban structure and micro object features. Evaluation using measurement data of 28 GHz in an urban environment shows that the proposed method can improve the modeling error by approximately 2.9 dB.

This paper proposes a phase shift method for reconfigurable meta-surface reflectors with failed passive elements, whose reflection phase cannot be controlled. To compensate for a decrease in the received signal power due to element failure, the proposed method adjusts the reflection phase by the same value for all passive elements from the initial reflection phase of each element, which is designed assuming a reflector does not fail. The proposed method provides higher in reflected signal power at the desired angle with less computational complexity than an individual adjustment of the reflection phase of each passive element. Numerical simulations and experimental results show that the proposed element failure correction achieves a higher received signal power than ignoring the element failure.

In Beyond 5G and 6G, the use of frequencies above 100 GHz is being considered to achieve ultra-high-speed and high-capacity communication. One application is to use these frequencies to connect peripheral devices such as wearable devices and user equipment (UE) to improve the communication speed of the uplink by using virtual terminal technology. To evaluate the wireless performance of this technology, it is necessary to clarify the radio wave propagation characteristics in the vicinity of the human body and construct a propagation channel model. However, the radio wave propagation characteristics of frequencies above 100 GHz in the vicinity of the human body have not been clearly elucidated, and estimation methods have not been established. Therefore, we have conducted radio wave propagation experiments, analyzed the propagation mechanism in the vicinity of the human body and examined estimation methods for propagation loss. In this paper, we will examine the shielding loss of the human body using radio wave propagation experiments in the 300 GHz band with a human body phantom. It has been shown that the loss due to human body shielding can be accurately estimated using a screen diffraction (SD) model, but there are limitations in accuracy above 100 GHz. Therefore, we propose estimation using the double isolated edge (DIE) model by approximating the human body with two screens. We demonstrate that this method can accurately estimate diffraction caused by the head and torso using evaluation with actual measurement values.

The Intelligent reflecting surface (IRS) has recently been attracting attention as a solution for coverage hole problems in mobile communication areas. Since the reflection range of the IRS depends on each direction and beamwidth of the IRS, the half-power beamwidth (HPBW) becomes an important parameter to reflect signals to user equipment widely distributed in the coverage hole. On the other hand, the reflection power decreases as the HPBW increases. Therefore, the design of the reflection pattern with appropriate HPBW is important to reflect sufficient signals to the desired range. Although the conventional reflection pattern design method, which divides the IRS into sub-IRSs with different reflection directions and synthesizes each reflection power, achieves a wide beamwidth of IRS, the reflection power is decreased due to the interference between the reflection power of each sub-IRS. In addition, the relationship between the HPBW and the reflection direction of each sub-IRS is not clarified. To solve this problem, we (1) propose the reflection phase optimization method to avoid the interference of each sub-IRS. We also (2) derive the relationship between the reflection HPBW and the gain of IRS. We (3) verify the validity of the proposed method through numerical analysis. In addition, we (4) verify that the proposed method achieves wide HPBW with high gain through experimental analysis. As the results show, the proposed method achieves about 1 dB reflection power enhancement. In addition, the proposed method can achieve HPBW up to 40 degrees of the IRS with the size of 50 wavelengths.

Wireless emulation technologies have recently been the subject of research and development to evaluate and verify wireless communication systems, including Beyond 5G, in a virtual space. To accurately emulate wireless system behavior, it is essential to model the radio propagation characteristics according to the environment. In recent years, several methods for site-specific path loss modeling have been proposed using machine learning, particularly convolutional neural networks (CNNs). Most of these methods simply stack convolution and pooling layers, but may not learn sufficiently due to vanishing gradients. Additionally, the most fundamental parameter characterizing propagation loss is the distance between the transmitter and receiver. However, some learning methods may fall into the local optimum solution by giving importance to other parameters. This paper proposes a method that uses Residual Network (ResNet) as the architecture and pre-training with Free Space Path Loss (FSPL) as the target variable. The proposed method improves learning efficiency and is expected to improve prediction accuracy. Finally, the effectiveness of the proposed method will be demonstrated by evaluating measurement data acquired in an urban area.

This article discusses the feasibility of classifying various situations for collaboration among medical staff engaged in surgery from videos taken in operating rooms. In video processing for medical use, especially for supporting surgery, many previous studies have attempted to recognize, from intraoperative videos, surgical phases constituting the workflows of specific surgical operations. However, it should also be significant to analyze, from videos obtained from observation cameras in operating rooms, how medical staff engaged in surgeries collaborate, with the aim of improving the efficiency and safety of collaboration among medical staff. The type of collaborative situations that occur are not predetermined but rather depend on actual behavior of medical staff during surgeries. Therefore, this study attempts to segment videos of surgical operating rooms into different collaborative situations by clustering the video frames based on similarities in the arrangements of medical staff. The arrangements are represented by the distributions of medical staff in the operating rooms as well as their observed postures observed in the video frames. In the results of our experiment with the video of a real surgery, some meaningful clusters of collaborative situations for medical staff could be obtained.

We propose a novel path loss estimation method based on deep learning with some newly defined system parameters and images. Estimating the radio wave propagation environment is one of the key techniques for indoor/outdoor high-speed wireless communication. The radio wave propagation environment is basically a multipath environment, and path loss characteristics should be estimated under various environments. The authors have already proposed path loss estimation methods based on machine learning and spatial image data. The purpose of this paper is to further enhance the path loss estimation accuracy by appropriately selecting the input parameters and the CNN/FNN model structure.

The virtualized terminal using a frequency-converting radio frequency (RF) relay device has been proposed as a new terminal configuration to achieve the 100+ Gbps uplink (UL) speeds required for Beyond 5G (B5G) mobile communication [1]. In B5G, where time division duplex (TDD) is considered to be applied as a duplexing method, the RF relay device should be able to detect UL and downlink (DL) timing and switch the relay circuit by the power fluctuation of the received signal for compact size and low power consumption. We propose a new timing detection method for RF relay devices that uses not only the received DL signal, but also the received UL signal. The lower bounds of the required carrier to noise ratio (C/N) and received power at the receiver to enable detection of TDD timing were obtained from the received signals for both UL and DL by simulation. Based on the experimental value of the conventional type of user equipment (UE), the receiver sensitivity of the UE communicating with the RF relay device using the 4.8 GHz frequency bandwidth was estimated. As a result, we found that even if the received DL signal of the RF relay device is lower than the lower limit of timing detection, it can still be received by the UE, and the RF relay device can switch the relay circuit if the UL signal from the UE has sufficient received power for timing detection, thus increasing the distance that can be communicated between the base station (BS) and UE.

Intelligent reflecting surfaces (IRSs) have recently been attracting attention as solutions for coverage hole problems in mobile communication areas. To utilize the IRS for mobile networks, the reflecting signal strength of IRS is evaluated in advance to estimate the improvement of coverage and interference with other wireless systems. In order to evaluate an actual IRS precisely, the reflecting phase and amplitude should be evaluated considering the mutual coupling effect. In addition, a large measurement system is required to satisfy the far-field conditions of IRS. This makes the evaluation of large IRS difficult. To solve this problem, we propose a practical evaluation method of a large IRS by synthesizing RCS patterns of small IRSs (sub-IRS). To verify the validity, we also formulate mutual coupling effects on the IRS. Through numerical evaluation, the influence of mutual coupling between each sub-IRS is reduced by increasing the size of the sub-IRS. In addition, through evaluation with the electromagnetic simulator, the RCS pattern of the proposed method reproduced that of a large IRS.

Wireless emulation is an essential technique for efficiently designing and verifying wireless systems in a virtual space. To accurately emulate the behavior of wireless systems in various scenarios, it is crucial to simulate radio propagation with high accuracy according to the environment. Several recent studies have proposed site-specific modeling methods using supervised machine learning. However, supervised learning generally requires a large amount of training data, and acquiring a large amount of measurement data on the propagation characteristics requires significant resources. Therefore, the model accuracy with few measurement data is an issue. This paper quantitatively clarifies the relationship between the amount of training data and the model accuracy. Furthermore, we propose a propagation modeling method with high accuracy with few data using fine-tuning. In this method, we pre-train the propagation model at a frequency with many measurement data available. Using pre-trained weights as initial values enables highly accurate propagation modeling with few data at a new frequency. Evaluation using measurement data in the 800 MHz, 2.1 GHz, and 3.5 GHz bands acquired in urban areas demonstrates the effectiveness of the proposed method in terms of accuracy and training time.

Beyond 5G/6G, terahertz and sub-terahertz bands are being considered to realize ultra-high-speed, high-capacity communications. As one of the use cases, virtualized terminal technology is being considered to improve the communication speed of the uplink by applying it to the connection between such peripheral devices as wearable devices and a user equipment and to use the high frequency for communication in the vicinity of the human body. In order to evaluate the wireless performance of this technology, it is necessary to elucidate the characteristics of radio propagation in the vicinity of the human body and construct a propagation channel model. However, the propagation characteristics of radio waves in the vicinity of the human body in these high-frequency bands have not been clarified, and a modeling method has not been established. In this paper, we analyze the propagation mechanism in the vicinity of the human body through radio propagation experiments in the 100 GHz band. Furthermore, we clarify the diffraction and scattering characteristics caused by the human body and the parameters necessary for modeling them and propose a modeling method for propagation loss that considers these characteristics. Finally, evaluation using actual measurements shows that the proposed model can estimate accurately with an RMSE of 4.40 dB.

We discuss how to extract building map images to be used in a learning-based method for predicting radio wave propagation. Learning-based prediction methods for radio wave propagation use building map images as spatial data, and we have already proposed a prediction method that uses images around Tx, Rx, and their midpoint. The method works well but has missing regions between Tx and Rx when the distance between them increases. In this paper, we further modify the method to include the whole area between Tx and Rx and also their surrounding regions in a single image. We present the method of using the dataset generated from measured data and an evaluation of how much the proposed method improves the prediction accuracy of radio wave propagation.

In order to realize a wireless emulator, a propagation model with high accuracy adapted for each specific environment is required. A stochastic propagation model is mainly used for the modelling of spatiotemporal propagation characteristics. However, a stochastic propagation model cannot predict accurately because no consideration is given to the surrounding environment. Therefore, a site-specific propagation model considering the surrounding environment is required to realize a wireless emulator. The prediction of multipaths is important for emulating of spatiotemporal parameters because spatiotemporal parameters are affected from multipaths. In particular, multipaths with one reflection or diffraction arriving via a wall or an edge of visible buildings from a transmitter or a receiver have a larger impact than other multipaths. Reflection and diffraction of these multipaths occurred on visible buildings from a transmitter or a receiver. In this paper, the site-specific propagation model for spatiotemporal parameters by a convolutional neural network (CNN) using a spherical image is proposed. The proposed model can predict spatiotemporal parameters because the spherical image just describes visible buildings from a transmitter and a receiver. It was clarified by evaluation that prediction by the proposed model can provide a correct description regarding both statistical characteristics in the evaluation area and specific characteristics at each point. We clarified the effectiveness of the proposed model by comparing it with the conventional model.

In higher frequency bands such as the millimeterwave band, which is expected to be utilized in the future, the received power fluctuates significantly due to changes in the propagation environment. Therefore, the spatial and temporal spectrum sharing between different wireless systems will be more practical. Some methods have been proposed to predict future spectrum usage, but prediction errors are unavoidable. There is a trade-off between the detection rate of future available spectrum resources and the false detection rate caused by prediction errors. Still, there is no established method for balancing these two factors. One conservative prediction to reduce false detections is to set a certain margin. However, the fluctuation tendency depends on the location, which results in an excessive margin in some sites, resulting in a lower detection rate. In this paper, we propose a method to set a location-dependent margin by considering the error characteristics of the constructed prediction model of received power. The proposed method enables us to increase the detection rate while reducing the false detection rate. The evaluation results by simulation in an urban area show that the proposed method can improve the detection rate by 10.6% while keeping the false detection rate, compared to when using a constant margin.

For the efficient use of limited frequency resources, the spectrum sharing allocation between mobile networks and different existing radio systems is promising. In this scheme, we focus on an allocation method that involves sharing spectrum bands with a mobile network operator (MNO). Recently, to improve the spectrum efficiency, dynamic spectrum allocation methods according to the MNO demand were proposed. However, in demand-based dynamic allocation, channel discontinuity in the frequency and time directions in the same MNO occurs when mapping the channel to MNOs. This channel discontinuity reduces the utility of shared frequencies. To address this problem, we propose a novel spectrum allocation method that maximizes frequency and time continuity for demand-based dynamic spectrum sharing. Our performance evaluation shows that compared with the existing methods, the proposed algorithm can improve the spectrum efficiency by 6%, channel continuity by 23%, and fairness by 5%.

This paper models the concept of figural goodness of a shape and apply the model to solving the visual occlusion problem. Based on the observation that 'goodness' of figure is influenced by local and global shape contexts, we model simplicity and regularity of a shape, and define a figural goodness score by linearly combining these two. To evaluate the usefulness and limitation of the proposed scoring, we applying the model to solving a contour completion problem. In the application, various completion paths of occluded contours are generated by a computer using different kinds of completion algorithms. To the generated paths, the best one is selected as the contour completion result by comparing their figural goodness scores. Experimental results have shown that the scoring model of figural goodness works effectively for the contour completion problem.

Research and development for the commercial deployment of 5G are accelerating and millimeter wave bands, which can utilize a wider bandwidth to realize larger capacity than conventional frequency bands, have become a focal point. These millimeter wave bands are different from conventional frequency bands in propagation characteristics, such as large blockage loss. Furthermore, the 5G system also assumes the case of new indoor use, such as industrial robots in factories and so forth. Therefore, a suitable propagation model for millimeter wave bands is necessary to evaluate 5G system performance by simulation or to design 5G area. In this paper, measurement campaigns of the path loss, the power delay profile and the azimuth angular profile of arrival were conducted in an indoor factory environment. In addition, we propose a suitable propagation estimation model for the indoor environment and we clarify that the proposed model is more suitable than the conventional model of the International Telecommunication Union Radiocommunication sector (ITU-R) in a subway platform by applying the measured results to a subway platform.

This paper presents a system for similarity retrieval of 3D models. In the system, a 3D sketch created with clay is used as a query. The clay sketch is digitized into a point cloud with a desktop 3D scanner. To compensate for the lack of scanning accuracy, the digitized model is refined by interpolation and decimation of data points. The effectiveness of the proposed system is experimentally evaluated by comparing the query-by-clay sketch method with other methods. The results have shown that the retrieval performance of the query-by-clay sketch method is superior to conventional methods.

The impression of a video changes depending on the audio and visual content of the video. Adding appropriate background music to a video is an important process for making the video more impressive. In this paper, we present a system called OtoPittan which recommends background music for video based on the valence and arousal model. As input for the system, first a user registers a video with no background music, and then inputs a desired impression by setting the valence and the arousal level. The system recommends music clips as candidates of background music for the video. Recommended music clips are determined by minimizing the difference between the desired impression and the predicted impression calculated from the audio features of each candidate music and the visual features of the registered video. We have confirmed that with the proposed system users can quickly create videos giving desired impressions.

In this paper, we propose an object decomposition system which can deal with partly occluded objects. In order to develop shape-independent object decomposition, the system combines different types of contour completion methods, namely, line completion, curve completion, line-symmetry completion and rotational symmetry completion. The system evaluates figural goodness of each completion pattern generated by these completion methods from the aspect of simplicity and symmetricity, and then selects the best one to output. From the experimental results, we have confirmed that the proposed system can correctly decompose objects from an input image.

Gaze writing (text entry by eye gaze) is an indispensable way of communication for motor-disabled people. While many interfaces for gaze writing are available today, the purpose of these interfaces is just a replacement of a keyboard. In order for users to communicate more efficiently, this paper introduces an interface for text copy and paste into a gaze writing system. The interface enables a user to specify a copy range and paste text to a desired position with only gaze interaction. To overcome the lack of accuracy of gaze pointing, the proposed system adopts a model of two step cursor control, which processes initial cursor setting and adjustment in turn. The experimental results have shown that the proposed interface makes gaze writing more efficiently.

In this paper, we propose a framework for extending existing matting methods to actualize more reliable alpha estimation. The key idea of the framework is integration of multiple alpha maps based on their reliabilities. In the proposed framework, the given input image is converted into multiple grayscale images having various luminance appearances. Then, alpha maps are generated corresponding to these grayscale images by utilizing an existing matting method. At the same time reliability maps (single channel images visualizing the reliabilities of the estimated alpha values) are generated. Finally, by combining alpha maps having the highest reliabilities in each local region, one reliable alpha map is generated. The experimental results have shown that reliable alpha estimation can be actualized by the proposed framework.

This paper reports improvement of indirect matching, which is a fast CBMIR (content-based music information retrieval) framework proposed in our previous study. Indirect matching achieves fast retrieval by combining offline search with representative queries and online quick similarity estimation based on the results of the offline search. We have found that the retrieval accuracy of indirect matching decreases when representative queries have little variation. This paper proposes a method for selecting representative queries having wide variation. To ensure wide variation between representative queries, the proposed method combines MDS (multi-dimensional scaling) and Ward's clustering. Experimental results have shown that the retrieval accuracy of indirect matching can be stabilized by the proposed method.

This paper considers the problem of estimating causal effects of a treatment on a response using supplementary variables. Under the assumption that a treatment is associated with a response through a univariate supplementary variable in the framework of linear regression models, Cox (1960) showed that the estimation accuracy of the regression coefficient of the treatment on the response in the single linear regression model can be improved by using the recursive linear regression model based on the supplementary variable from the viewpoint of the asymptotic variance. However, such assumptions may not hold in many practical situations. In this paper, we consider the situation where a treatment is associated with a response through a set of supplementary variables in both linear and discrete models. Then, we show that the estimation accuracy of the causal effect can be improved by using the supplementary variables. Different from Cox (1960), the results of this paper are derived without the assumption of Gaussian error terms in linear models or dichotomous variables in discrete models. The results of this paper help us to obtain the reliable evaluation of causal effects from observed data.

This paper presents workload evaluation of gaze-writing systems. As a method for evaluating user's workload of a gaze-writing system, we used NASA-TLX, a questionnaire-based method for workload evaluation. We partly modified the NASA-TLX questionnaire for users of gaze-writing systems. We experimentally compared the workloads of users of two traditional gaze-writing systems QWERTY and Dasher. The experimental results showed that the workload of QWERTY tended to be lower than Dasher.

This paper proposes a method for contour completion of a skew-symmetry object which is partly occluded by other objects. The proposed method assumes that skew symmetry of a shape can be measured by correlation with its reflection on a no orthogonal plane. Under the assumption, the method estimates a skew-symmetry axis and a skew transverse angle, and then reflects the contour to the estimated axis considering the skew-transverse angle. Using the reflected contour as a patch, the occluded part of the contour is restored. However, due to the information loss caused by occlusion, the reflected contour slightly moves from the correct position. The method moves the reflected contour within a small range to smoothly connect it with the original contour. We experimentally confirmed the effectiveness and limitation of the proposed method by comparing the proposed method with the human perception.

In traditional heterogeneous cellular networks (Het-Net), base stations (BS) basically associate users based on received signal power. However, in the case of multiband HetNet, in which macro BSs and smallcell BSs use different frequency bands, this conventional cell association method is not effective because there is no interference between macro BSs and smallcell BSs. Additionally, there are huge differences in coverage and available bandwidth. These differences cause inefficient association which causes the imbalance between achievable rate and traffic demand. In order to overcome this problem, we propose a novel cell association method based on the combinatorial optimization. The proposed method considers achievable rates, traffic demands and the number of users belonging to each BSs simultaneously. Numerical simulation results show that the proposed association method achieves system rate gain twice as high as the conventional one.

This paper proposes a collaborative filtering based spot recommendation system seamlessly available in user's home areas and away areas. The system recommends spots to a user which are highly evaluated by other users whose preference in spot genre selection is similar to the user's one and whose home areas cover the user's current location. As an application of the proposed system, we have implemented a restaurant recommendation system. Experimental results have shown that the superiority of the proposed system to an existing spot recommendation system.

This paper presents a method for extracting regions of non-rigid moving objects from a video. The proposed method utilizes interactive image cutout and local affine transformation. In the proposed method, once a user draws seed-lines on the object and background regions of the first frame of a video, the seed-lines are automatically tracked to next frames based on local affine transformation, and the object regions in each frame are automatically extracted by image cutout using the tracked seed-lines. Experimental results have shown that the proposed method can correctly extract the regions of nonrigid objects from a video.

This paper proposes a system for completing the contour of an object which is partly occluded by other objects. The proposed system classifies objects into rotational symmetry objects, line symmetry objects and the other objects. To a rotational symmetry object, the system estimates the angle and center of rotation of the object focusing on the periodicity of the curvature change of the contour. Rotating the object with the estimated angle and center of rotation, the system reconstructs the occluded part of the contour. To a line symmetry object, the system detects line symmetry axes. The occluded part is completed by folding the shape to the line symmetry axes. To other shapes, the system uses a spline interpolation to complete the contour. We have experimentally confirmed the effectiveness and limitations of the proposed method by comparing the proposed method with human perception.

This paper presents a fast similarity retrieval method for vector images. To reduce the computational cost of similarity matching, the proposed method uses pre-calculation results of similarity matching, which are obtained in advance by matching DB images with previously selected images called representative queries. At runtime the proposed method just matches the actual query (the user-inputted query) and the representative queries. Comparing the similarities with the precalculated similarities, the proposed method quickly estimates the actual similarities of DB images to the actual query. Experimental results have shown that the retrieval time is greatly reduced by the proposed method without much deterioration of retrieval accuracy.

Adding BGM (background music) to a video is an important process in video creation because BGM determines the impression of the video. We model impression estimation of a video as mappping from computer-mesurable audio and visual features to impression degrees. As an application of impression estimation of a video, we propose OtoPittan, a system for recommending BGM for helping users to make impressive videos. OtoPittan regards the problem of selecting BGM from a music collection as a partial inverse problem of the impression estimation. That is, to an inputted video and desired impression, BGM which produces a good match to the desired impression when adding it to the inputted video is recommended. As implementation ways of impression estimation of a video, we use a static user model and a dynamic user model. The first model statically constructs a mapping function learnt from training data. The second model dynamically optimizes a mapping function through user interaction. Experimental results have shown that the static user model has high estimation accuracy and the dynamic user model can efficiently performs optimization without much user interaction.

A novel method for content-based 2D vector image retrieval is proposed in this paper. To retrieve vector images, first, the proposed method extracts WFPs (Weighted Feature Points) from vector images. In this extraction process, the method directly accesses the parameters of Bezier curves defining the shapes of vector images and computes WFPs of the images according to these parameters. Second, to a pair of vector images, the method evaluates similarity by matching WFPs using PTD (Proportional Transfer Distance). In the proposed method, vector images are not rasterized, which enables real-time retrieval. As a prototyping, we have implemented a system for sketch-based vector image retrieval. In the system, a sketch is used as a query. To bridge the representation gap between sketches and vector images in a database, the strokes in sketches are approximated with Bezier curves as a pre-process. Using the system and a database containing 1,120 logomarks, we have confirmed the retrieval performance and have discussed the limits of the proposed method.