Poster Award

Yuto Mori and Nobuyuki Umezu (Ibaraki University, Japan)

We propose an interactive physical book constructed from Japanese traditional handmade paper, onto which local folklore narratives are projected. A book remains one of the most fundamental interfaces for even very young children, offering intuitive navigation for storytelling playback. The system locates the current page using QR codes unique to each sheet and detects the user's touch via capacitive sensors embedded beneath the paper surface. Four sets of picture shows were selected from a local folklore archive in a museum. Users first choose characters to paint; these characters subsequently appear in the projected story, thereby enhancing engagement. We conducted a series of user studies with visitors to a science museum. Thirty-eight participants interacted with the prototype and then completed a five-point Likert-scale questionnaire. The results yielded high scores for interest (Mean=4.45), usability (4.42), and comparison with conventional books (3.79). These findings suggest that participants perceived advantages in both traditional books and our interactive projection book.

Chung Ko Han and Yoke-Hou Tan (Chung Yuan Christian University, Taiwan)

Domain-specific question answering with large language models (LLMs) faces two critical barriers: commercial API costs that scale linearly with query volume and data-privacy risks from transmitting sensitive corpora to external services. We present a signal-adaptive Graph RAG architecture for Traditional Chinese Bible QA that addresses both challenges through intelligent retrieval routing. Our system integrates three heterogeneous databases - PostgreSQL (100,222 structured records), Qdrant (3,041 dense embedding vectors), and Neo4j (19,317 nodes, 58,031 relationships) - orchestrated by a priority-ordered decision-tree router. A signal detector extracts six boolean query features (verse reference, chapter reference, multi-book scope, multi-person mention, event keyword, and place name), then selects among six specialized retrieval routes, each combining a tailored subset of SQL, semantic search, graph traversal, and cross-reference engines. We evaluate the architecture on 100 curated Traditional Chinese questions spanning five categories (verse lookup, topic, person, event, and general) using 19 metrics across retrieval, LLM-judged generation, and semantic similarity dimensions. Comparing Claude Haiku 4.5 (commercial API) against Gemma 3 4B (locally deployed via Ollama at zero API cost), we find that 6 of 7 generation metrics differ by fewer than 4 percentage points, with complementary strengths: Claude leads on faithfulness (0.887 vs. 0.849) while Gemma leads on answer relevancy (0.658 vs. 0.592). Route-dependent analysis reveals that direct SQL lookup achieves the highest accuracy (0.841-0.886) whereas the most complex cross-reference route scores lowest, suggesting that graph density is a key bottleneck. These results demonstrate that a well-designed signal-adaptive retrieval architecture can substantially reduce the impact of model scale, enabling cost-free, privacy-preserving deployment with competitive generation quality.

Elton Alencar (Federal University of Amazonas UFAM, Brazil & Venturus - Innovation and Technology Center MAO, Brazil); Edma Urtiga de Mattos (Federal University of Amazonas, Brazil); Diego A. Amoedo (Universidade Federal do Amazonas & Agência Nacional de Telecomunicações - Anatel, Brazil); Victoria Guimaraes, Pedro V S Matias, Emanuel Oliveira and Luca Naja (UFAM, Brazil); Matheus Figueiredo (Federal University of Amazonas, Brazil); Pedro Oliveira, Matheus Uchoa, Cristian Maia, Lucas Botinelly, Rosiane Brito, Bruno Cardoso, Laura Leite and Lucas Pessoa (UFAM, Brazil); Ana Gabriela Pires (Federal University of Amazonas, Brazil); Alexandre Miranda (FPF Tech, Brazil); Agemilson Pimentel, Ruan Belem and Rômulo Fabrício, Jr. (TPV Technology, Brazil); Celso Barbosa Carvalho (Federal University of Amazonas, Brazil); Waldir Silva (Universidade Federal do Amazonas, Brazil)

Consumer electronics manuals remain a source for configuration, use, and troubleshooting, but present unique challenges for automated question answering due to their specialized symbols, procedural diagrams, and domain-specific terminology. Generalist Vision-Language Models (VLMs) struggle to accurately interpret such content without domain adaptation. This work presents a parameter-efficient for technical Visual Question Answering (VQA) in consumer electronics manuals, emphasizing the trade-off between cross-product generalization and product-specific specialization. A synthetic VQA dataset (2264 question-answer pairs from 295 manual pages) was constructed using DeepSeek-OCR. LoRA-based adaptation was evaluated across five VLM backbones. The results demonstrated performance improvements: Qwen3-VL-8B achieved an Exact Match of 23.84% on multi-product data. Single-product specialization on monitor manuals reached 43.59%, demonstrating that product-specific adaptation reduces generic responses. All datasets, training scripts, and evaluation artifacts are publicly available in our GitHub repository: https://github.com/eltonalencar90/vlms-user-manuals-vqa.

Ming-Lin Chuang (National Penghu University of Science and Technology, Taiwan); Kuan-Ming Wu (National Penghu University of Science & Technology, Taiwan); Shu-Min Tsai (National Penghu University of Science and Technology, Taiwan)

This study presents a concurrent rectifier with a dual-band matching network for wireless power transfer. The proposed design reduces reflection and improves RF-to-DC conversion efficiency at two operating frequencies. Experimental results show that simultaneous dual-frequency input produces higher DC output voltage and better efficiency than a single-frequency input with twice the power.

Hideya So (Kogakuin University, Japan); Jin Nakazato (Tokyo University of Science, Japan)

Unmanned aerial vehicle (UAV)-assisted relay networks are expected to provide temporary communication infrastructure in disaster areas. As a method for determining UAV deployment, cooperative control using multi-agent reinforcement learning (MARL) has been investigated. However, existing learning frameworks do not consider radio propagation or inter-link interference. This study evaluates the signal to interference plus noise ratio (SINR) characteristics of UAV deployments. Numerical analysis shows that the SINR degrades when simultaneously transmitting UAVs are aligned on the same straight line. Moreover, narrowing the beam directivity improves the SINR, whereas in the linear placement, narrowing the beam directivity does not effectively improve the SINR performance.

Katsuori Takeuchi (University of Electro-Communications, Japan); Ryo Yamamoto (The University of Electro-Communications, Japan); Nobuhiko Itoh (Shibaura Institute of Technology, Japan); Satoshi Ohzahata (The University of Electro-Communications, Japan)

A hybrid Packet Loss Concealment (PLC) method combining linear predictive coding and deep neural networks is effective for real-time audio transmission. However, most previous research focuses on single-instrument sound sources, and their application to ensemble sound sources has not been fully explored. This paper proposes a hybrid PLC method incorporating frequency information processing based on spectral structures, aiming for auditorily natural and high-precision restoration of ensemble sound sources. Comparative experiments with previous research demonstrate the effectiveness of the proposed method.