Published papers

This paper proposes a fast beam tracking algorithm for wireless backhauling to moving vehicles having access point function. The basic concept is developed earlier as the adaptive moving network to track to spatio-temporal traffic demand fluctuation. This architecture necessitates wireless connection between vehicular access points and network side. Further, wireless backhaul requires high capacity as well as coverage, therefore use of millimeter wave and massive array beamforming is essential. Our backhauling strategy focuses on the movable region of vehicles on the road to simplify the beam tracking algorithm. Simulative evaluation verifies its capability even when the vehicle speed dynamically changes.

This paper experimentally demonstrates 512 color shift keying (CSK) signal transmission for optical camera communication (OCC). We achieved error-free operation with a CMOS image sensor module and a multi-label classification neural network-based equalizer.

To achieve cost-effective mobile fronthaul (MFH) with a time-division-multiplexing passive optical network (TDM-PON), we present a prototype optical-mobile cooperation interface, which converts mobile upstream scheduling information to transmission request information for the mobile upstream data that will arrive at each optical network unit (ONU). By basing bandwidth allocation on the request information converted by the interface, low-latency upstream transmission can be achieved on the MFH with the TDM-PON. We demonstrate the time spent for the conversion process in the prototype interface under various and practical conditions for the first time.

We investigate the performance of a hybrid transmission system using a 2×PAM4 to 16QAM optical aggregation and 16QAM to 2×QPSK electrical disaggregation as flexible data center interconnects.

We investigate the performance of a hybrid transmission system using a 2×PAM4 to 16QAM optical aggregation and 16QAM to 2×QPSK electrical disaggregation as flexible data center interconnects.

We investigate the performance of a hybrid transmission system using a 2⨯PAM4 to 16QAM optical aggregation and 16QAM to 2×QPSK electrical disaggregation as flexible data center interconnects.

This paper presents a real-time implementation of deep joint source-channel coding (DeepJSCC) for image transmission over 5G networks using a field-programmable gate array (FPGA). DeepJSCC simultaneously performs source and channel coding using a single neural network and is known for its robustness in low-SNR environments. However, deploying such large-scale models on GPUs often leads to high power consumption and jitter, making real-time processing on edge devices challenging. To address this issue, we design and implement a convolutional neural network (CNN)-based DeepJSCC model on an FPGA evaluation board using INT8 quantization and high-level synthesis (HLS). The proposed architecture achieves low latency and high throughput while maintaining moderate resource utilization. We integrate the FPGA with a commercial 5G base station (gNB) and user equipment (UE) through modified I/Q interfaces. The experiment confirms that the FPGA-based DeepJSCC system consistently reconstructs images at a frame rate of approximately 33 FPS with stable peak signal-to-noise ratio (PSNR) performance.

In recent years, access and metro networks have been transitioning toward architectures based on disaggregation and virtualization, where software-based implementation of physical-layer signal processing on general-purpose computing platforms has become an important technical challenge. In next-generation access systems, digital coherent transmission enables high-order modulation and polarization multiplexing but requires computationally intensive digital signal processing, such as polarization demultiplexing and adaptive equalization, to be executed in real-time. However, conventional time-domain equalization suffers from computational complexity that increases linearly with the number of equalizer taps, making real-time operation challenging under long-tap conditions caused by chromatic dispersion compensation and future wideband transceivers. To address this issue, this paper proposes a multi-thread frequency-domain adaptive equalizer (MT-FDE) that fully exploits the parallel processing capability of a graphics processing unit (GPU). By combining fast Fourier transform (FFT)-based frequency-domain processing with batch processing, the proposed method reduces computational complexity to logarithmic order with respect to the tap length and enables real-time adaptive equalization including polarization demultiplexing. Numerical simulations and offline and real-time experiments demonstrate that the proposed MT-FDE improves the tolerance to increased tap numbers by a factor of 4.8 and achieves real-time operation under long-tap conditions while maintaining transmission performance comparable to that of conventional time-domain equalizers, providing practical design guidelines for fully virtualized coherent access networks.

As one of the realization for real-time remote monitoring and object recognition using high-definition camera images that can support safe and automated driving, this paper proposes image size reduction by differentiation from the background, focusing on fixedly installed cameras at Road-Side Unit (RSU). The video images acquired by the camera installed on RSU are transferred to the edge server, and information related to the traffic situation is recognized by image processing. Since it requires high-capacity transmission, the use of the millimeter-wave band having large bandwidth available is essential. Meanwhile, a multi-hop relay is desirable due to its short coverage. In this case, a multi-hop relay should support video image traffic from multiple RSU nodes and hence it accumulates the transmission latency. The proposed scheme greatly reduces the size of the image that needs to be transmitted, while maintaining object (vehicle) detection accuracy.

This paper proposes a value-aware spatial stream allocation strategy for deep joint source-channel coding (DeepJSCC) incorporated with multiple-input multiple-output (MIMO), aimed at enhancing image transmission quality under extreme environment such as underwater scenarios. In MIMO eigenmode transmission, each stream can achieve a different diversity combining gain, with the principal stream exhibiting the highest gain. The proposed scheme allocates spatial streams of individual eigenmodes to image blocks based on their perceptual importance, which is extracted using a saliency map. Simulation results confirm that the proposed scheme achieves higher perceptual quality and structural similarity, especially in low-SNR conditions. This work demonstrates the potential of integrating semantic awareness into physical layer design, contributing to robust and efficient image communication in mission-critical applications.

This paper proposes image transmission using multiple-input multiple-output (MIMO) Deep Joint Source-Channel Coding (DeepJSCC) with focusing on value-based spatial stream allocation. DeepJSCC integrates source and channel coding through deep learning, enabling robust image transmission even under low signal-to-noise power tratio (SNR) conditions. For applications such as underwater acoustic communication, limited communication capacity due to significant SNR degradation over long distances is a challenging issue. Key proposal of this paper is to assign signal stream of MIMO eigenmode spatial division multiplexing (E-SDM) for divided image blocks according to their visibility importance. Its effectiveness is validated in terms of structural similarity index measure (SSIM) and edge preservation index (EPI) under low SNR regime.

Visible light communication(VLC) is currently attracting attention toward the realization of 6th generation mobile communication systems (6G). Optical camera communication (OCC), which is one of the realizations of VLC, can be constructed using general-purpose commercial devices such as display or LEDs as the transmitter and optical camera as the receiver. It can be seen as a promising alternative of electromagnetic-wave-based wireless communications especially in underwater or aerial communications to extend network coverage. OCC has variety of challenges to expand its application region: communication coverage, capacity, and stability. In this paper, we propose a method to optimize imaging parameters such as exposure time and gain for crosstalk compensation in the nonlinear region of the image-sensor inputs to improve the communication quality of OCC.

This paper evaluates the impact of accuracy in a parallel eigenvalue calculation method on transmission characteristics. We report that parallel computing does not affect the characteristics when parameters are drastically changed.

This paper evaluates the impact of accuracy in a parallel eigenvalue calculation method on transmission characteristics. We report that parallel computing does not affect the characteristics when parameters are drastically changed.