https://ir.lib.ncu.edu.tw/handle/987654321/59 Search the Channel s https://ir.lib.ncu.edu.tw/simple-search https://ir.lib.ncu.edu.tw/handle/987654321/99388 title: 基於遮蔽生成式Transformer的風格生成表達;基於遮蔽生成式Transformer的風格生成表達 abstract: 本研究利用一種數據驅動的方法，用於從中性表情生成個人化的微笑風格圖像，目標在於產生多樣化的微笑風格，同時保留個人的臉部特徵。相較於傳統需要耗費大量人力進行臉部屬性標註和合成的生成模型不同，我們的方法先利用表情分類器(Emotion classifier)搭配GradCAM[1]來自動提取與身份無關的臉部表情注意力區域，這些關鍵區域接著用於引導一個遮蔽生成式 Transformer（Masked Generative Transformer）。實現更具表情一致性與個人特徵保留的圖像生成。 整體架構包含兩個階段:在第一階段，我們採用VQGAN[2]來提取圖像的潛在空間標記(latent tokens)，同時融合表情分類器與GradCAM[1]產生的臉部表情注意力圖。透過這些來引導Transformer重建潛在標記，以生成初步的微笑圖像。 第二階段，專注於超解析度（super-resolution），我們將更高解析度的圖像輸入到另一個VQGAN[2]提取高解析度的潛在空間標記。這些高解析度潛在空間標記，同時融入前一個階段的低解析度標記，還有表情注意力圖。此階段的Transformer將整合第一階段生成的低解析度標記及高解析度標記資訊，以還原高品質及細節豐富的最終圖像。 實驗結果顯示，我們的方法在保留個人面部特徵的同時，能有效產生多樣且自然的微笑風格圖像，展現其於個性化表情生成領域的潛力。;This paper presents a data-driven approach for generating personalized smiling facial images from a single neutral expression. The objective is to synthesize diverse smile styles while preserving individual facial characteristics. In contrast to conventional generative models that rely heavily on manual annotation of facial attributes and synthesis procedures, the proposed method utilizes an emotion classifier combined with Grad-CAM to automatically extract identity-independent facial expression attention regions. These regions are subsequently used to guide a Masked Generative Transformer, enabling expression-consistent and identity-preserving image generation. The proposed framework comprises two stages. In the first stage, VQGAN is employed to extract latent tokens from the input image. Expression attention maps, derived from the emotion classifier and Grad-CAM, are integrated to guide a Transformer in reconstructing the masked tokens and generating an initial smiling image. In the second stage, which addresses super-resolution, a higher-resolution version of the image is processed through another VQGAN to extract high-resolution latent tokens. These are combined with the low-resolution tokens from the first stage and the expression attention maps. A second Transformer then reconstructs the final high-quality image by fusing both low-resolution and high-resolution information. Experimental results show that the proposed method effectively generates diverse and natural smile styles while preserving individual facial identity, demonstrating its potential in the domain of personalized facial expression synthesis. <br> https://ir.lib.ncu.edu.tw/handle/987654321/99386 title: 結合特徵交互作用分析與可信度約束優化之多階 段製程參數區間生成研究;AStudy on Multi-Stage Process Parameter Interval Generation via Feature Interaction Analysis and Credibility-Constrained Optimization abstract: 隨著工業4.0 的推進，多階段製造程序(Multistage Manufacturing Processes, MMPs) 的穩定性與良率優化已成為製造業的核心挑戰。MMPs具有高度耦合與參數交互作用的特性，任一階段的微小偏差皆可能引發連鎖性瑕疵。以紡織製程為例，整經(warping)、上漿(sizing)、併軸 (beaming) 及織布 (weaving) 等連續步驟間的參數波動，常導致斷紗(brokenwarp) 或間歇性瑕疵(intermittent warp)，造成嚴重的原料浪費與生產成本增加。 現有研究廣泛採用樹狀模型(Tree-basedModels)進行缺陷預測，但在實務應用上仍面臨兩大核心限制：首先，傳統模型生成的參數建議區間往往過於寬泛，缺乏精確的操作指導性；其次，單一決策路徑難以捕捉多參數間的交互作用(InteractionEffect)，導致優化成效受限。 為此，本研究提出一套整合預測模型、交互作用分析與受約束優化之參數區間生成架構。首先，針對關鍵瑕疵建立多目標預測模型fi，並導入ShapleyResidual等工具尋找交互作用的特徵間的交互關係。接著，推論特徵及其交互作用模式，建立參數優化規則。最後，本研究設計一組受約束的優化目標函數，在極大化效益指標(Estimated Benefit, EB) 的同時，加入規則長度(NumberofRules)作為可信度的條件，確保生成的規則R具備統計意義並有效避免過擬合。 實驗結果顯示，本研究所提方法能有效收斂參數區間，在提升效益基準(EB)的同時，顯著降低了製程中的瑕疵率。此系統不僅提供了具備可解釋性的製程建議，更為製造現場提供了具備操作彈性的區間控制策略，達成了從缺陷預測轉向「缺陷預防」的零缺陷製造目標。;With the advancement of Industry 4.0, optimizing stability and yield in Multistage Manufacturing Processes (MMPs) has become a core challenge for the manufacturing industry. Characterized by high coupling and complex parameter interactions, even minor deviations in any stage of an MMP can trigger cascading defects. In the context of textile manufacturing, fluctuations in process parameters across sequential steps—such as warping, sizing, beaming, and weaving—frequently lead to defects like broken warp or intermittent warp, resulting in significant material waste and increased production costs. Although existing research extensively employs Tree-based models for defect prediction, their practical application remains hindered by two primary limitations. First, the parameter recommendations generated by traditional models are often overly broad, lacking the precision required for operational guidance. Second, single-path decision rules fail to capture the interaction effects among multiple parameters, thereby limiting the effectiveness of optimization. To address these issues, this study proposes an integrated framework that combines predictive modeling, Feature Interaction analysis, and constrained optimization for generating precise parameter interval recommendations. First, multi-objective predictive models fi are developed for four key defect types, utilizing tools such as SHAPLEY RESIDUAL to extract core parameters influencing defects. Next, the patterns of individual features and their interactions are inferred to establish optimization rules. Finally, a constrained objective function is designed to maximize the Estimated Benefit (EB) while incorporating a Number of Parameters as a credibility constraint. This ensures that the generated rules R possess statistical significance and effectively mitigate the risk of overfitting. Experimental results demonstrate that the proposed method effectively streamlines the parameter selection process, balancing rule simplicity with high effectiveness. This significantly reduces defect rates while enhancing EB. By offering flexible interval-based control strategies, the system bridges the gap between defect prediction and prevention, advancing the goal of zero-defect manufacturing. <br> https://ir.lib.ncu.edu.tw/handle/987654321/99383 title: 結合MCP 與大型語言模型之互動式網路協定教學 系統;An MCP-Enabled Interactive Network Protocol Teaching System with Large Language Models abstract: 網路協定是資訊工程與資訊安全領域的基礎知識，然而當前的教學模式多依賴靜 態教材與單向講授，缺乏互動性與即時回饋，導致學生難以將抽象的協定理論與實際 封包數據建立有效連結。教師也面臨缺乏系統化框架輔助設計互動式課程，以及難以 即時掌握學生理解狀況的挑戰。此外，Wireshark等專業工具雖然能夠顯示封包的詳 細細節與封包資訊統計，但其介面複雜且學習曲線陡峭，增加初學者的學習門檻。 本研究提出一個結合大型語言模型(Large Language Models, LLM)、模型上下文協 定(Model Context Protocol, MCP) 與WebShark封包分析工具的智慧網路協定教學 平台。系統以Claude 4 Sonnet作為核心的理解與教學引導引擎，透過MCP 技術整 合WebShark 進行實際封包分析，實現語意理解與實務操作的結合。透過學生提問來引 導學生，並根據學生問題動態調整教學策略，提供個別化的學習體驗。 為評估系統成效，本研究採用G-Eval評估框架，使用GPT-4o與Gemini 作為雙評估器，針對Answer Relevancy(答案相關性)、Hallucination(幻覺程度)、Role Adherence(角色遵循) 與Knowledge Retention(知識保留) 四個維度進行評估。實驗設計 涵蓋HTTP、TCP、DNS 與HTTPS 等四種網路協定教學場景，比較本系統與純LLM 基 準系統的表現。 實驗結果顯示，本系統在雙評估器的評估下均展現顯著優勢。在需要具體數據包 證據的教學場景中，本系統的Answer Relevancy 顯著優於純LLM(GPT-4o vs 純LLM： 0.70 vs 0.36；Gemini vs 純LLM：0.79 vs 0.55)，證實MCP 工具整合能提供更相關且精 確的回應。Role Adherence 方面，本系統在兩個評估器下均維持高分(GPT-4o 約0.78， Gemini 達1.0)，而純LLM 在所有實驗中均為0，突顯純LLM 無法維持教學助理角色的 可能存在缺陷。本系統在Knowledge Retention 與Conversation Completeness 兩項指標均 達到優異表現，反映出系統能在多輪互動中維持脈絡一致性。值得注意的是，Gemini 在純LLM 的部分實驗中偵測到Misuse 現象，而GPT-4o 並未發現此問題，顯示雙評估 器策略能更全面地揭示系統潛在風險。 本研究的主要貢獻在於：(1) 設計並實作整合 LLM、MCP 與 WebShark 的混合式網路協定教學系統原型；(2) 透過對比實驗驗證工具輔助對系統 效能的提升，包括回應相關性、角色遵循、知識保留等多個維度；(3) 建立基於 G-Eval 雙評估者的多維度評估方法，為教學型 AI 系統效能 評估提供可行框架；(4) 提供可立即應用的學習輔助工具原型，降低 初學者使用 Wireshark 的學習門檻。;Network protocols constitute the foundational knowledge of Computer Science and Information Security. However, current pedagogical models rely heavily on static materials and oneway lectures, lacking interactivity and real-time feedback. This results in students struggling to bridge the gap between abstract protocol theories and actual packet data. Educators also face challenges such as a lack of systematic frameworks for designing interactive curricula and difficulty in monitoring student understanding in real-time. Furthermore, while professional tools like Wireshark provide comprehensive packet details and statistics, their complex interfaces and steep learning curves present significant barriers for beginners. This study proposes an intelligent network protocol teaching platform that integrates Large Language Models (LLMs), the Model Context Protocol (MCP), and the WebShark packet analysis tool. The system employs Claude 4 Sonnet as the core engine for understanding and pedagogical guidance. Through MCP technology, the platform integrates WebShark to perform live packet analysis, achieving a fusion of semantic understanding and practical operation. The system guides students through inquiry-based learning and dynamically adjusts teaching strategies based on student questions to provide a personalized learning experience. To evaluate the system’s effectiveness, this study utilizes the G-Eval framework, employing GPT-4o and Gemini as dual-evaluators. The evaluation focuses on four dimensions: Answer Relevancy, Hallucination, Role Adherence, and Knowledge Retention. The experimental design covers four protocol teaching scenarios—HTTP, TCP, DNS, and HTTPS— comparing the proposed system against a pure LLM baseline. Experimental results demonstrate that the proposed system exhibits significant advantages across both evaluators. In scenarios requiring specific packet evidence, the Answer Relevancy of this system was significantly superior to the pure LLM (GPT-4o: 0.70 vs. 0.36; Gemini: 0.79vs. 0.55), confirming that MCP tool integration provides more relevant and precise responses. Regarding Role Adherence, the system maintained high scores (approx. 0.78 by GPT-4o and 1.0 by Gemini), whereas the pure LLM scored 0 across all experiments, highlighting the baseline’s inability to maintain a teaching assistant persona. The system also achieved excellent performance in Knowledge Retention and Conversation Completeness, reflecting its ability to maintain contextual consistency over multiple interactions. Notably, Gemini detected ”Misuse” in some pure LLM experiments that GPT-4o failed to identify, underscoring the importance of a dual-evaluator strategy in revealing potential system risks. The main contributions of this research are: (1) designing and implementing a hybrid network protocol teaching system that integrates LLM, MCP, and WebShark; (2) validating the performance improvement through comparative experiments across multiple dimensions including answer relevancy, role adherence, and knowledge retention; (3) establishing a multi-dimensional evaluation framework based on G-Eval with dual evaluators, providing a feasible approach for assessing teaching-oriented AI systems; (4) providing an immediately applicable learning assistance tool prototype that lowers the learning barrier for beginners using Wireshark. <br> https://ir.lib.ncu.edu.tw/handle/987654321/99381 title: 用於物件偵測之輕量化局部全域特徵整合網路;LoGIN: A Lightweight Local-Global Integration Feature Pyramid Network for Object Detection abstract: 在多尺度物件偵測的應用情景中，由於特徵金字塔網路 (FPN) 本身的尺度離散特性，往往導致尺度斷層 (scale truncations) 以及特徵傳遞過程中的資訊衰減 (information attenuation)，使得偵測效能面臨許多挑戰。儘管近期的方法試圖透過複雜的特徵聚合機制或從層級結構上的密集堆疊來緩解這些問題，但它們往往會帶來高昂的運算成本，或是忽略了跨尺度特徵表示的連續性。為了解決這個問題，我們提出了基於輕量化設計建構的局部-全域整合網路 (Local-Global Integration Network, LoGIN)，有別於現有的方法，LoGIN 從特徵表徵連續性 (representational continuity) 的視角來解決尺度斷層問題。它透過更全面的特徵融合方法去擴展每個離散層級的有效尺度覆蓋範圍，來消除位於尺度邊界處物件的模糊性；同時利用全域上下文資訊來強化語義一致性，有效抑制背景雜訊。在 MS-COCO 資料集上的實驗結果顯示，LoGIN 在輕量級偵測器中展現了極具競爭力的效能，為多尺度物件偵測的實際應用提供了一個更穩健的解決方案。;Efficient multi-scale object detection faces significant challenges due to the discrete nature of Feature Pyramid Networks (FPNs), which often results in scale truncations and information attenuation during transmission. While recent approaches have attempted to mitigate these issues via complex aggregation mechanisms or dense layer stacking, they often incur high computational overhead or overlook the continuity of feature representation across scales. To address this, we propose the Local-Global Integration Network (LoGIN), constructed from a lightweight design. Distinct from existing methods, LoGIN tackles the scale gap problem through a perspective of representational continuity. It expands the effective scale coverage of each discrete level to resolve ambiguities for objects at scale boundaries, while simultaneously leveraging global context to enforce semantic consistency and effectively suppress background noise. Experimental results on MS-COCO demonstrate that LoGIN achieves competitive performance among lightweight detectors, providing a robust solution for practical applications. <br>