Best Paper Competition

Best Paper Competition

Best Paper Candidate
01 Vision-Language Augmented Multimodal Framework for Real-Time Violence Recognition
Syu Jhih Jhang and Tzu-Chia Huang (Tamkang University, Taiwan); Wan-Chi Yang (National Taipei University of Nursing and Health Sciences, Taiwan); Zhen-An Zhong and Chih-Yung Chang (Tamkang University, Taiwan)

A multimodal violence detection system is proposed, combining Siamese I3D, Dual-Stream 3D CNN, and BLIP-2+BERT for real-time surveillance. Dual-modality inputs (RGB + TV-L1 optical flow) feed two visual branches; a vision-language module provides semantic verification. On a 2,846-clip UCF Crime dataset, the Dual-Stream 3D CNN achieves 88.26% accuracy and 89.92% F1; the full multimodal system reaches 91.64% F1. An ablation study confirms the necessity of each component.
02 Analysis of ISO-Ring Connections on Robustness in ISO-NLDMOS
Shih-Chieh Chien (Vanguard International Semiconductor, Taiwan); Jian-Hsing Lee (Vanguard International Semiconductor Corp. & Vis Micro Inc., Taiwan); Chun-Chu Liu (Vanguard International Semiconductor Corporation, Taiwan); Yu-Sheng Chiu and Chieh-Yao Chuang (VIS, Taiwan); Chih-Fang Huang (National Tsing Hua University, Taiwan)

This work investigates the impact of isolated-ring (ISO-ring) connection schemes on the second breakdown current (It2) of a fully isolated laterally diffused NMOSFET (ISO-LDNMOS). The device demonstrates a significantly enhanced ESD robustness when the N-ISO is tied to the drain and the P-ISO is grounded, compared with configurations where both rings are grounded or floating. TCAD simulations reveal that grounding or floating both ISO-rings triggers a base push-out (kirk) effect after snapback, leading to premature device failure. In contrast, connecting the N-ISO to the drain while grounding the P-ISO effectively suppresses the Kirk effect, thereby improving It2
03 Methods for Improving the Efficiency of Terahertz Leaf Water Content Detection Using Resolution Compensation Technology
Ming-An Chung, Zhang Jun-Hao, Min Chiun Hsieh, Kai Xiang Chen, Chia Chun Hsu and Chia-Wei Lin (National Taipei University of Technology, Taiwan)

Terahertz (THz) imaging is widely used in non destructive measurement fields such as agricultural monitoring, biomedical testing, and materials analysis due to its high sensitivity to moisture. Changes in the water content of plant leaves are closely related to their physiological state. This study establishes a non-destructive detection system for leaf water content using a THz light source and an array camera. The distribution of water within the leaf was observed by exploiting the fact that water strongly absorbs electromagnetic radiation. This study addresses the insufficient resolution of raw THz imaging employing bicubic interpolation combined with edge detection to enhance structural contrast. Experimental results show that the processed images can more clearly present the leaf vein outline and moisture gradient, thus improving the interpretability of THz images.
04 Code Rate Improvement of Armknecht HE through Packing
Jin Nakashima and Haruhiko Kaneko (Institute of Science Tokyo, Japan)

Homomorphic encryption (HE) has attracted attention as a privacy protection technology for cloud computing and consumer electronics. One such scheme, proposed by Armknecht, provides high-speed operations (excluding setup) while suffers from a low code rate. In this paper, we propose a packing scheme that extends the message support from a single point to a set of multiple points. This approach improves code rate proportionally to the number of packed elements while maintaining the same level of security. Our evaluation results demonstrate that the processing time per unit of data can be reduced to a maximum of approximately 1/390.
05 Dual-Branch-Based Distributed Blind Estimation Over Wireless Sensor Networks
Sulin Chi (Otemon Gakuin University, Japan); Hsuan Fu Wang (National Formosa University, Taiwan); Tetsuya Shimamura (Saitama University, Japan)

Distributed in-network processing in wireless sensor networks (WSNs) has attracted considerable attention for estimating the unknown parameter of interest across various scientific and engineering disciplines. However, conventional distributed estimation algorithms typically require prior knowledge of the training signal or the desired signal. To overcome this limitation, distributed blind estimation methods, such as distributed blind equalization (d-BE), have been developed to recover the transmitted data signal without the prior knowledge of the training signal. The performance of the d-BE is easily affected by the transmission channel. To address this issue, dual-branch-based d-BE approach is proposed. Computer simulation results demonstrate that the proposed method achieves significant performance improvements over conventional methods.
06 ECG-PPG Fusion with Feature Transfer for Blood Pressure Estimation
Liang-Hung Wang (Fuzhou University, China)

Accurate cuffless blood pressure (BP) monitoring is crucial yet vulnerable to noise and domain shifts. We propose a feature-transfer framework that fuses raw ECG and PPG with discriminative features from BP tri-class ResNet and ECG binary DenseNet, and regresses Systolic Blood Pressure/Diastolic Blood Pressure (SBP/DBP) using ResNet blocks and a BiLSTM in a multitask network. On a public dataset, the model achieves MAE±SD of 3.04±4.93 mmHg (SBP) and 1.76±3.37 mmHg (DBP), meeting AAMI and BHS grade A. After transfer to ICU data, performance remains compliant, supporting real-world deployment.
Best Student Paper Candidate
01 Hybrid Multiple Access Scheme Based on Interference-Aware Switching Between MU-MIMO and Band-Splitting NOMA-MRT
Shogo Kurimoto and Daichi Kishino (Waseda University, Japan); Keita Kuriyama (NTT Corporation, Japan); Fumiaki Maehara (Waseda University, Japan)

A hybrid multiple access scheme is proposed that dynamically switches between multi-user multiple-input multiple-output (MU-MIMO) and band-splitting non-orthogonal multiple access with maximum ratio transmission (NOMA-MRT), based on user-side interference sensing. To mitigate inter-cell interference, which is a critical issue in densely deployed cellular networks, each user equipment estimates the interference level by measuring the received power during signal-absent periods and reports it to the base station. The base station selects MU-MIMO under low interference conditions to exploit spatial multiplexing, and switches to band-splitting NOMA-MRT under high interference, where frequency subband partitioning reduces intra-cell interference. Simulation results demonstrate that the proposed scheme achieves robust throughput performance across various interference conditions, outperforming traditional MU-MIMO, NOMA-MRT, and orthogonal multiple access (OMA-MRT) schemes.
02 Hybrid FEM-Machine Learning Optimization of a Multilayer SPR Biosensor for Enhanced Cervical Cancer Detection
Rithika Sivasankar and Aishwarya Mol S (Vellore Institute of Technology, Chennai, India); Reya Pradeep (SRMIST, India); Mohana M and Reena Monica P (Vellore Institute of Technology, Chennai, India)

This paper presents a machine learning-assisted optimization framework for enhancing the performance of a multilayer Surface Plasmon Resonance (SPR) biosensor for cervical cancer detection. Reflectance characteristics were generated using Finite Element Method (FEM) simulations in COMSOL Multiphysics for a BK7 prism-based structure operating at 633 nm. Normal (n = 1.368) and cancerous (n = 1.392) analyte conditions were modeled to extract resonance angle shift, sensitivity, and Figure of Merit (FOM). The baseline configurations achieved an average sensitivity of 276.54 °/RIU and a mean FOM of 3673.49. A Random Forest regression model was trained using multilayer thickness parameters to predict and optimize FOM. The optimized configuration achieved a predicted FOM of 5501.99, corresponding to a 49.8% improvement over baseline structures. The results demonstrate that integrating FEM-based simulation data with machine learning enables efficient and scalable optimization of high-performance SPR biosensors for early-stage cervical cancer detection.
03 Adaptive Transfer Learning Strategies for a Vision Transformer-Based Seismic Foundation Model
Ahmed Alghaili (Universitas Islam Indonesia, Indonesia); Isack Farady (Yuan Ze University, Taiwan); Chih-Yang Lin (National Central University, Taiwan); Ming-Jen Wang (National Applied Research Laboratories, Taiwan)

The interpretation and analysis of seismic signals using deep learning have gained significant attention in recent years. In particular, pretrained foundation models have emerged as a promising approach in seismic imaging, offering improved generalization capability and reduced dependence on large labeled datasets. However, the effectiveness of different transfer learning strategies for downstream seismic tasks remains insufficiently explored. This paper presents a comparative study of frozen-encoder and fine-tuning adaptation strategies applied to a Vision Transformer-based Seismic Foundation Model (SFM). Using a unified experimental protocol, we evaluate both strategies on seismic reconstruction and interpretation tasks. The results demonstrate that fine-tuning provides superior performance for interpretation-oriented applications, while frozen-encoder adaptation achieves competitive results for reconstruction tasks with substantially lower computational cost. These findings offer practical insights and guidelines for the efficient deployment of pretrained seismic transformer models in real-world applications.
04 Development of a Virtual Process PDK for KLayout with ChatGPT as an Auxiliary Tool
Chia-Hung Hsu (National Taipei University of Technology, Taiwan); Yue-Fang Kuo (Yuan Ze University, Taiwan); Ming-An Chung (National Taipei University of Technology, Taiwan)

This paper presents a high-efficiency and cost-effective workflow for Process Design Kit (PDK) development assisted by ChatGPT. Ruby-based Parameterized Cell (PCell) templates are automatically generated using ChatGPT and extended to support multiple technology nodes through parameterized design rules. A virtual 180-nm MOS PCell template is first developed and subsequently scaled to 90-nm and 65-nm processes, from which both PMOS and NMOS devices are derived by modifying the well and implantation layers. In addition, the Design Rule Check (DRC) rule file is optimized to enable seamless switching among different processes. The generated PCells are verified through Design Rule Check (DRC) and Layout Versus Schematic (LVS) checks. The proposed approach significantly reduces development time and manpower, demonstrating strong potential for scalable and efficient PDK development.
05 Noise Analysis of Laser Diode Using Fourier Transform and Phasor Diagram
Yuki Kazama and Mitsuru Shinagawa (Hosei University, Japan); Soukichi Funazaki, Jun Katsuyama, Yoshinori Matsumoto and Shinichiro Tezuka (Yokogawa Electric Corporation, Japan)

Noise waveform reconstruction of a laser diode was studied for an application of an electro-optic sensor system to low-frequency measurement. The noise waveform can be reconstructed by inverse Fourier transform from the modeled power and phase spectra based on measured noise power and phase spectra. The reconstructed noise waveform is close to a measured noise waveform. It was confirmed that the noise waveform reconstruction method is proper.
06 Edge-Aware Dual-Path Spatial-Temporal Graph Attention Networks for Traffic Flow Forecasting
Shu-Jian Gao, Tzu-Chia Huang, Hsiang-Chuan Chang, Sheng-Yi Ding and Chih-Yung Chang (Tamkang University, Taiwan)

Accurate traffic flow forecasting is essential for intelligent transportation systems, yet real-world road networks exhibit both explicit connectivity constraints and latent correlations induced by shared temporal patterns and dynamic road conditions. We present an edge-aware dual-path spatial-temporal graph attention framework for multi-step traffic forecasting. The model employs two parallel spatial paths, combining a road-topology graph with an adaptive graph to capture hidden dependencies beyond direct connectivity. Unlike node-only attention, our spatial module injects edge attributes-static road length and time-varying road speed-directly into attention scoring and message aggregation, enabling condition-aware propagation under dynamic link states. Temporal dependencies are modeled by a dilated Temporal Convolutional Network (TCN) with residual connections to capture both short-term bursts and longer-range trends. A gated fusion module adaptively integrates the two paths. The proposed design retains the inductive advantages of attention-based graph learning while improving expressiveness through edge-aware interactions, making it suitable for realistic traffic systems where link conditions strongly modulate regional dynamics.