DSpace collection: 學位論文

Digital Image Recovery and Multiple-watermarking Techniques

Wed, 15 Apr 2026 06:12:01 GMT

title: Digital Image Recovery and Multiple-watermarking Techniques abstract: The research topic of this paper is to integrate the digital image processing schemes and the watermarking techniques, and those methods will apply on the digital images and digital videos. The research topic includes three parts: (1) image recovery, colorization and enhancement, (2) multiple-watermarking techniques, and (3) the integration of image recovery and multiple-watermarking techniques. The abstracts of all chapters are described below: Chapter 1 – Image Recovery The lacuna texture synthesis is proposed for the virtual restoration of ancient Chinese paintings and digital images. Lacuna texture synthesis is a patching method, which uses the Markov Random Field (MRF) model. We eliminate the undesirable patterns, such as stains, crevices, and artifacts, and the algorithm fills the lacuna regions with the appropriate textures. The proposed scheme not only maintains a complete shape, but also prevents the edge disconnection in the final results. Chapter 2 – Visible Watermark Removal In this chapter, an image recovery algorithm for removing visible watermarks is presented. Independent component analysis (ICA) is utilized to separate source images from watermarked and reference images. Three independent component analysis approaches and five different visible watermarking methods are examined in our study. The experimental results will show that visible watermarks are successfully removed, and that the proposed algorithm is independent of both the adopted ICA approach and the visible watermarking method. Moreover, several watermarked images sourced from various websites are removed the watermarks successively. Chapter 3 – Image Colorization In the past, the artists adopted the black ink to represent various sights and objects in Chinese ink-and-wash, such as, mountain scenery, waterscape, animals, plants, etc. This chapter will introduce an effective method to colorize the Chinese ink-and-wash paintings. The proposed method not only takes fewer computing time than the conventional method, but it also can preserve the soft-gradual tone in the ink-and-wash paintings, such as, water-flowing, smog, cloud, waterfall, and shadow etc. Chapter 4 – Image Enhancement We will introduce the weighted histogram separation (WHS) in this chapter, which is presented to enhance the high dynamic range images. The property of weighted histogram separation situates between the successive mean quantization transform and the histogram equalization. Additionally, the proposed method is further applied to the local enhancement, which is termed as the adaptive weighted histogram separation (AWHS). Chapter 5 – Spatial Domain Multiple-watermarking Algorithm The objective of our study in information security is to develop a multiple watermarks embedding and extraction algorithm, which is called as spatial domain multiple-watermarking algorithm. This algorithm is one kind of quantization index modulation, it can impose bi-watermark or tri-watermark on the host image. Furthermore, the extracted watermarks not only are exploited to detect the tampered areas, but it is also used for attack classification and attack identification. Chapter 6 – Dual Domain Bi-watermarking Algorithm A dual domain bi-watermarking algorithm embeds bi-watermark into the host image in discrete-cosine-transform domain (DCT), and it is the extension of the spatial domain bi-watermarking algorithm. However, the bi-watermark can be extracted from both spatial domain and DCT domain. By the same token, two separated watermarks from the extracted bi-watermark have different capability for various compression rates, and they also reveal the different robustness against the global and the regional attacks. Chapter 7 – 2.5 Domain Tri-watermarking Algorithm In this chapter, we will introduce an integration of dual domain bi-watermarking algorithm and visual cryptography, which is named as 2.5 domain tri-watermarking algorithm (2.5D-TW). This algorithm implements tri-watermark embedding in discrete-cosine-transform domain (DCT) for video protection, but the tri-watermark can be extracted from both spatial domain and DCT domain. Three separated watermarks from the extracted tri-watermark reveal the different robustness against various attacks. According to the bit error rates of those three watermarks, the algorithm even identifies whether the attack is occurred in spatial domain or in temporal domain for video. Chapter 8 – Integration of Image Recovery and Watermarking Algorithm The key of this chapter is to integrate the image recovery scheme and the watermarking technique. The spatial domain bi-watermarking algorithm is used to add the halftone of downscaled host image into the host image. After extracting the bi-watermark from the covered image, the bi-watermark is restored to the gray-scale image using the proposed inverse halftoning, which utilizes the linear programming and quadratic programming. Furthermore, the bi-watermark is not only exploited to detect the tampered areas without prior data, but it also can be applied to recover the tampered areas in the tampered image. Chapter 9 – Linear Programming and Its Applications In 1736, the great mathematician Leonhard Euler published a paper to solve the problem of seven bridges of Königsberg, and he translated it into the graph theory problem. This study is the well-known Euler circuit problem. Here, we solve non-Euler circuit problem using mix-integer linear programming, which transforms the non-Euler circuit to the Euler one. In addition, the binary integer programming is exploited to determine the edge direction. The experimental results will show that the proposed scheme can be applied to the route planning, the continuous line drawing and the real-object production. Chapter 10 – Conclusions and Future Works Consequently, we will summarize the previous researches and describe the possible improvement and applications in the future works.

結合多目標追蹤與人物再辨識之影像式異常偵測系統

Thu, 16 Oct 2025 06:50:58 GMT

title: 結合多目標追蹤與人物再辨識之影像式異常偵測系統

在無線感測網路中最小化路徑長度及最大化覆蓋之充電技術

Thu, 16 Oct 2025 05:59:59 GMT

title: 在無線感測網路中最小化路徑長度及最大化覆蓋之充電技術 abstract: 近年來，隨著無線感測網路技術的快速發展和物聯網應用的普及，無線感測裝置的充電技術也愈來愈受到重視。如何設計合適的充電機制來提供感測器運作之電量，以延長無線感測網路的生命週期已成為現今最熱門研究主題之一。然而，現存大多數的充電機制，其基本概念是由一台具移動力的充電車，分別移動至每個感測器的可充電範圍內，再逐一地對各個感測器進行充電，導致充電車在執行充電任務時所需移動的充電路徑長度，將隨著感測器數量增多而有顯著的增加，進而造成充電車需要花費大量的時間與電量在充電的移動過程中。另一方面，現存的充電機制大多設計讓充電車移動至感測器充電範圍內的某定點後，停留在該定點並執行充電任務，待感測器充電完成後，充電車才繼續移動前往下一個目標感測器之充電範圍內的定點停留，並執行充電任務，換句話說，充電車的移動過程會是走走停停的狀態，無法維持等速移動，導致充電車必須花費更多的電量來執行充電任務，因此也降低了充電效率。有鑑於此，本論文提出兩種不同的充電技術，分別為Recharging Path Construction (RPC) 技術與Coverage Aware Energy Replenish Mechanism (CAERM) 技術，用以改善現存充電機制的效能。首先，本論文所提出之 RPC 技術，其探討充電車在等速移動下，同步進行充電工作時，如何規劃最佳的充電路徑，進而讓充電車的移動方式更有效率。接著，本論文所提出之 CAERM 技術，其將感測器之充電優先權納入考量，針對不同位置之感測器，分析其覆蓋面積對於整體感測的貢獻度，並協助充電車動態規劃其充電路徑，以進一步改善充電效率。實驗結果顯示，本論文所提出的兩種充電技術，可有效的解決現存充電機制所產生的問題，大幅提昇充電車執行充電任務的效能。

永續環境願景：發掘人工智慧的潛力以建構智慧生態系統

Thu, 16 Oct 2025 05:52:05 GMT

title: 永續環境願景：發掘人工智慧的潛力以建構智慧生態系統 abstract: 台灣的水、能源、材料和勞動資源受到氣候變遷的影響，因此，確保環境永續性對於未來世代的福祉至關重要。本研究提出了一種全面應對綜合性環境系統的方法，需要在宏觀、中觀和微觀層面實施政策變革。為追求與自然環境的和諧共生中，本研究重視環境管理、智慧農業的實踐，並積極探索創新的菌絲體策略。本研究所提到的環境系統中，包括四個關鍵方面：地下水位、PM2.5空氣污染嚴重程度、氣候等問題之預測及提昇生成牛樟芝的效益。在地下水位預測中，混合的人工智慧(AI)模型在R2和RMSE指標方面均超過基準，其高精確度使得其成為決策者潛在的參考依據，以進行高效的地下水資源規劃。對於PM2.5預測，混合AI模型由於額外的ACT輸入而顯著提升性能，對於準確的區域預測至關重要。該混合AI模型對於公眾意識和減少污染的政策實施是不可缺少的工具。混合AI模型在基於中央氣象局(CWB)數據生成精確氣象預測方面表現出色，更減少對物聯網設備的依賴。這些氣象預測提供於農民將有助於溫室微氣候預測和智慧溫室的建構。該模型萃取高維數據的特徵能力强大，使其能夠準確預測農試所(TARI)和伸港(Shengang)溫室的微氣候趨勢的變化。此外，混合AI模型有助於提取顯著特徵，可靠地估計菌絲體產量並實現條件的人工智慧優化，展示了提高生產的巨大潛力，成功減少了75％的時間消耗(相比於基準)。透過協同整治宏觀、中觀和微觀的環境系統，使得台灣更有潛力進一步朝向永續發展目標（SDGs），有效地對抗資源匱乏的威脅。此方法適用於應對眼前的環境挑戰，為未來奠定永續社會的基礎。

多模態感測器融合物件偵測技術於自動駕駛之研究

Wed, 12 Mar 2025 06:56:57 GMT

title: 多模態感測器融合物件偵測技術於自動駕駛之研究

Exploiting Token Constraint and Multi-Scale Memory Bank of Contrastive Learning Based Vision Transformer for Unsupervised Person Re-identification

Wed, 12 Mar 2025 06:47:30 GMT

title: Exploiting Token Constraint and Multi-Scale Memory Bank of Contrastive Learning Based Vision Transformer for Unsupervised Person Re-identification abstract: 本論文提出了視覺變換器 (ViT)字符約束和多尺度記憶庫（TCMM）方法，以解決現有最先進的無監督行人再識別工作中遇到的區塊雜訊和特徵不一致問題。對比學習的快速發展在無監督行人再識別任務中取得了顯著成功。許多先前優秀的無監督行人再識別方法生成偽標籤、計算分群原型、使用 ViT 提取特徵，最後通過對比學習來訓練模型。偽標籤方法在領域適應和無監督學習問題中已顯示出有希望的結果。然而，ViT 通過首先執行區塊嵌入來處理影像，這不可避免地會引入區塊中的雜訊，甚至可能包括不同的身份實例，從而損害再識別模型的效能。另一方面，現有的偽標籤方法經常丟棄難以分群的離群樣本，這犧牲了離群樣本的潛在價值，導致模型的多樣性和穩健性有限。為了解決這些問題，本論文引入了 ViT 字符約束來限制 ViT 的輸出字符特徵，以減輕區塊雜訊對 ViT 架構造成的損害。此外，提出的多尺度記憶庫通過樣本級和原型級樣本增強了模型對離群樣本的探索，並保持了特徵的一致性。本論文還整合了我們先前提出的生成對抗網路和修補生成模型，提供了額外的正樣本以增強模型的多樣性。本論文提出了一種基於生成對抗網路和修補模型的額外對比正樣本生成和選擇策略，以從另一個角度探討無監督行人再識別任務。實驗結果表明，我們的系統在常見的行人再識別資料集基準上達到了最先進的效能。

Sim2Real 在高動態環境下人形機器人的平衡控制

Wed, 13 Mar 2024 06:16:33 GMT

title: Sim2Real 在高動態環境下人形機器人的平衡控制 abstract: This thesis presents comprehensive research into the dynamic balance control of a humanoid robot, namely the Robinion2S. The research initiates with detail of humanoid robot platforms with mechatronic systems, walking gait algorithms, and perception systems. Special focus is given to Robinion2S, the latest version of the Robinion series, which forms the backbone of the study. The experiments show the limitations of traditional PID-based balance control methods when deployed in a complex, dynamic environment such as a balance board. Despite optimization efforts using a high-throughput random search algorithm within Nvidia’s Isaac Gym simulation environment, the PID-based approach fails to ensure consistent balance. This result leads to the need for more robust control strategies. The research focuses on reinforcement learning techniques to balance control to overcome the result. Despite the challenges of traditional control theory, reinforcement learning techniques show potential as a viable solution to the intricacies of balance control. The reinforcement learning models demonstrate their adaptability and robustness in maintaining balance, hinting at their potential to solve more complex control problems. Extending the study into real-world applications, the Sim2Real approach is developed. The Sim2Real approach implements the trained reinforcement learning models into a dynamic, physical environment. Despite not achieving ideal results, the approach demonstrates the potential for trained models to transfer control policies effectively from simulation to the real environment. This thesis provides potential methods in the field of balance control in humanoid robots, motivating a shift from traditional control methods to more robust reinforcement learning techniques. Despite not being ideal, the obtained results show significant potential for future research and advancements in the robotics field. This research provides a foundation for understanding the balance control in the humanoid robot and potential strategies to optimize the performance in a real-world environment.

基於深度學習類神經網路之機器人動作決策認知系統

Fri, 06 Oct 2023 06:22:53 GMT

title: 基於深度學習類神經網路之機器人動作決策認知系統 abstract: High-dimensional complex motion generation is an interesting research topic. Most action generation methods in robotics research use a single pose as the model output. However, in some scenarios, only a series of motions can be output at one time. The calligraphy writing task belongs to a complex motion generation challenge which needs to output a series of motions at one time. The calligraphy writing task can be divided into position learning and posture learning. For position learning, human can directly form a properly rational statement of where to write. In Taylor’s problem categories, the position learning problem in calligraphy learning belongs to Q3 and Q4 types which are formal statement. That is, human can easily design an algorithm to generate a policy to robot. In the contrast, humans are not able to describe the relationship between the writing posture and the writing result. Therefore, the posture learning problem in calligraphy learning belongs to Q1 and Q2 types in Taylor's problem categories. In order to solve the problems of Q1 and Q2, this dissertation will propose the fundamental cognitive system with self-learning ability. This dissertation integrates the framework of human perception, memory, and decision-making into the robot system through the cognitive psychology. We use the top-down and bottom-up processing of the human perceptual system to design a perception model of the cognitive system, which enables encoder networks to learn online. In the memory model, we implement the psychological multi-store model with a deep neural network, so that robots can remember past events like humans. We use the hypothesis generation model of psychology in the decision-making model, so that the robot has a human-like thinking process. Integrating these cognitive models, robots can generate action strategies based on their goals through their own experience. Finally, we use a practical robot as experimental platform to verify the learning ability of the proposed cognitive system.

使用機器學習演算法建構預測存活以及費用模型 -以冠狀動脈繞道手術病患為例

Fri, 06 Oct 2023 06:19:55 GMT

title: 使用機器學習演算法建構預測存活以及費用模型 -以冠狀動脈繞道手術病患為例 abstract: 冠狀動脈繞道手術（CABG）對冠狀動脈疾病（CAD）患者是一種有用的治療方法。在一些相關研究中，潛在疾病和合併症會影響死亡率和再入院率，這些問題將直接增加醫療費用。但是，目前沒有研究找出會影響存活以及醫療費用的危險因子。本研究使用國家健康保險研究數據庫（NHIRD），健保資料庫是全台灣最大且最完整的資料庫，包含各種醫學信息、病患的門診、住院資訊。本研究中選擇首次接受CABG 手術的患者，並使用不同的機器學習演算法(LGR, LR, CART, MARS, RF, SVR, XGBoost)透過特徵選取，找出影響存活以及費用的相關危險因子，再進行預測以及評估。本研究結果顯示透過特徵篩選有助於提高模型預測率、準確性且可以利用較少的危險因子進行預測。CABG 患者共病患有腎臟疾病是影響存活的關鍵因子且腎臟疾病的病患醫療費用較高; 術前一年醫療費用、當次手術費用以及洗腎的次數是預測術後一年費用的關鍵因子。本研究可以幫助政府制定良好的醫療政策，並可以朝著準確的預防性醫療、降低醫療總費用，及更有效的醫療管理的方向發展。 Coronary artery bypass surgery grafting (CABG) is a useful treatment for patients with coronary artery disease (CAD). In some related studies, underlying disease and comorbidities will affect mortality and readmission that will directly increase medical expenses. However, the researches currently barely focus on figuring out the most important risk factor that may affect the survival rate and medical expense of the CAGB patients This study used the National Health Insurance Research Database (NHIRD), which is the largest dataset that includes comprehensive medical information, and inpatients and outpatients diagnose records in Taiwan. We have selected patients who received their CABG surgery for the first time and have used different machine-learning algorithms, including: LGR, LR, CART, MARS, RF, SVR, and XGBoost for selecting the most important variables, then evaluated and predicted the major factors that impact survival rate and medical expenditure. From this study, it shows that feature selection can help improve the prediction capabilities and accuracy of the models with fewer factors. The result indicated that kidney disease that CABG patients have is a risk factor that affect survival rate and increase medical expenses. The medical expenditure in one-year before surgery, surgical expense, and the times of HD were strong predictors of future expense. This research can help the government to formulate medical policies better. Moreover, it enhances preventive medical prediction more precisely, reduces the total medical expenses possibly, and develops more effective medical management.

Memetic Computing for Automatic Music Composition

Thu, 13 Oct 2022 02:40:44 GMT

title: Memetic Computing for Automatic Music Composition abstract: Music exerts a ubiquitous influence on human cultures and daily lives. Composing music is deemed rather complicated because it involves various factors (e.g., instruments, melodies, rhythm, and chords) needed to be well coordinated for creating harmony, tension, and emotions. This study focuses on an important domain of evolutionary composition methods, i.e., rule-based evolutionary systems. By including music knowledge, this type of systems is able to compose quality music fully automatically. Many of these systems applied weighted rules for music quality evaluation. The elaborately crafted rules and weights power the evolutionary composition systems, yet unintentionally limits extendibility of the systems. Based on the existing studies, we reorganized the music knowledge involved and proposed the deparameterized evaluation function (DEF) to improve the flexibility and transparency of rule-based evaluation. As a case study, the DEF is applied to composing bossa nova music. Our evolutionary system for bossa nova is satisfactory in music quality and acceptable in efficiency. The transparency of the DEF allows human's validation, and furthermore, the development of advanced memetic algorithms. Evolutionary multitasking (EMT), an influential paradigm of memetic computing, features the ability to solve multiple optimization tasks in parallel. A state-of-the-art EMT algorithm is the multifactorial evolutionary algorithm (MFEA). Despite of its remarkable performance, several issues are found impeding the search. This study accordingly proposed an advanced version named the multifactorial evolutionary algorithm with resource reallocation (MFEARR), which adaptively alters the searching behavior to improve the usage of evaluation resources. Tested on the multitask optimization (MTO) benchmark suite, the proposed MFEARR shows significant improvement in both efficiency and effectiveness. Moreover, the MFEARR possesses the potential to improve the traditional evolutionary composition systems in composing polyphonic (multi-melody) music because different melody composition tasks usually share some common ground in evaluation criteria. In other words, the tasks are mutually correlated and presumed to facilitate one another through EMT. To exploit this advantage, a novel MTO problem is formulated for polyphonic music composition and the evolutionary multitask composer (EMTC) is developed to solve this problem. The empirical results support the effectiveness of the MTO problem formulation and the efficiency of the EMTC in composing polyphonic music. 音樂在文化與日常生活中，有著無所不在的影響力。音樂創作因為需同時考量諸多要素而被視為相當複雜的工作。此研究著重在演化式作曲的一個重要的領域──基於規則評估的演化系統。這類系統藉由納入音樂知識來自動地創作優質的音樂作品。許多這樣的系統採用具權重的規則來評估音樂品質。雖然那般精心設計的規則和權重能驅動演化作曲系統，然而卻限制了系統的使用彈性。基於現存的研究，我們重新組織相關的音樂知識來設計一個去參數化之評估函數，以改善基於規則評估的透明性與彈性。此去參數化之評估函數亦被應用於自動創作bossa nova音樂作為一個案例分析。我們的演化系統可以有效率地創作令人滿意的bossa nova音樂作品。去參數化之評估函數的透明性允取使用者進一步為其開發進階的瀰因演算法 (memetic algorithm)。作為瀰因計算中頗具影響力的領域，演化多工 (evolutionary multitasking) 能平行化地解決多個最佳化工作。其中一個最先進的演化多工演算法為「多因演化演算法」 (multifactorial evolutionary algorithm)。儘管它擁有出色的效能，它仍有些許問題可能對搜尋過程帶來不利的影響。此研究因此提出「具資源再分配之多因演化算法」 (multifactorial evolutionary algorithm with resource reallocation) 作為一個改良版本。這個改良版本能因應搜尋當下狀況自動調節搜尋行為來改善評估資源的利用率。「具資源再分配之多因演化算法」在一系列的演化多工的效能測試問題上，都能顯著地提升搜尋效能。除此之外，在創作多聲部音樂的問題中，它還具有改良傳統演化作曲系統的潛力。這類問題中包含多個旋律創作工作，而由於這些工作具有某種程度的關聯性，所以可以透過演化多工來促進工作之間的相輔相成。為了利用這個優勢，我們針對多聲部音樂作曲制訂了一個新穎的多工最佳化問題，並開發了「演化多工作曲系統」 (evolutionary multitask composer) 來解決這樣的多工最佳化問題。實驗結果驗證了這個問題制訂的有效性，以及演化多工作曲系統的作曲效率。