電腦科學與資訊工程

臺灣的許多人要使用電腦打字時皆會選擇「注音輸入法」作為輸入的方法。然而要使用注音輸入法輸入中文時若無切換輸入法則有可能會誤用到英文輸入法。如要輸入「今天天氣好」五字，使用英文輸入法輸入時會輸出「rup wu0 wu0 fu4cl3」。這種文字不易理解。本研究的目的即為研究將未切換到注音輸入法而打出的英數符號混和字元翻譯為漢字的方法。 本研究使用「PTT 中文語料」與「維基百科中文資料庫」訓練 GRU、BiGRU、LSTM 和 Transformer 以及計算維特比演算法，並與 Google 輸入工具的進行比較。整體來看以 PTT 中文語料計算的維特比演算法的 BLEU4 分數最高，在準確率以及 BLEU4 的評分皆高於Google 輸入工具，分數分別為 0.94 與 88.3 分。 本研究之成果在應用方面極廣，可應用於線上翻譯或聊天軟體的即時翻譯。本研究使用之程式碼開源於 GitHub 頁面，除了可讓使用者下載使用外，使用者也可訓練自己的模型。

目前網路上流傳許多使用人工智慧修復照片的網站或應用軟體。然而,由於這些訓練資料多數來自國外,導致修復中式建築照片的效果欠佳。此外,許多老舊照片因氧化、潮濕而泛黃,使得修復程序比起修復純黑白相片更加困難。因此,本研究旨在建🖂一個專門修復中式建築物的機器學習模型,主要分為以下三個部分:首先,使用機器學習模型對老舊照片進行修復,包括著色、去模糊化和降噪;其次,分析使用不同比例之有色調照片(模擬泛黃照片)訓練模型的效果;最後,研究不同的修復順序(著色、去模糊化、降噪)和模型執行次數對照片修復效果的影響,發現「著色、去噪、去模糊化」的順序修復效果最佳。此外，許多老舊照片因為受損等原因，只剩下極少的特徵，因此本研究採用機器學習模型，以延伸重建原始照片。透過這種方法，我們能夠重新建構當時建築物周圍可能的場景和情境。

本研究提出了一種基於毫米波FMCW雷達的即時居家健康監測系統，與RGB攝影機感測器方案相比，該系統更具成本效益並保護用戶隱私。對於居家健康監測系統來說，除了走、站、坐、臥四種常見的身體動作外，及時發現緊急情況也至關重要。因此，我們支持跌倒偵測以及兩種手勢的識別，一種用於頭痛，另一種用於緊急事件。然而，由於雷達點雲的稀疏和不規律特性，我們提出一種新的身體動作波形表示方式和一套處理程序用來平滑動作波形，利用雷達點雲的三度空間座標和速度在時間序列上的變化波形來表示動作特徵並作為神經網路的輸入，再搭配一輕量級的1D-CNN+LSTM神經網路來實現即時動作辨識。根據實驗結果，此方法可以達到30FPS的輸出效率和94%的辨識準確率。

在高度數位化的社會，存在大量的數位影像資料，在不影響視覺品質之前提下如何標記持有者或資料來源是一重要課題，而不可視浮水印機制是可行的解決之道，本研究運用對抗性機器學習技術的概念及深度學習技術研發設計數位影像之浮水印機制，研製偵測器與註冊器。本研究設計研製之偵測器與註冊器可以處理任意大小的影像，經實驗分析具高度的保真性，並具可以承受 JPEG 中度品質(qlt=50)的失真壓縮攻擊，解壓縮還原之已嵌浮水印影像偵測器仍舊可以有效判定具浮水印。本機制可以結合網站、伺服器或影像設備為其提供的數位影像嵌入浮水印，在不影響視覺感官的前提下標記來源或持有者。

本研究旨在探索大型語言模型如何應用於音樂生成故事。研究動機源自音樂作為文化中不可或缺的一部分，但若要以文字精準表達出音樂中的故事情緒尚屬困難，藉由本研究提出的方法可以使故事顧及到音樂的情緒起伏。隨著 AI 的發展，我 們開始看到它們在各領域的應用。這項研究的目的是製作出一個系統能以音樂作為輸入，輸出音樂內的故事，為達成目的，我們結合多個模型。研究使用 PyTorch等工具，並探討文句和音樂的共同表示方法，實現情感匹配。研究結果顯示，音樂和文句情感辨識模型表現不錯，也研發出一個完整的生成流程。目前已有直接生成音樂的模型，也有把音樂統整介紹的模型，卻沒有依據音樂中的情緒生成故事的模型。我們研究就是在解決這個問題，結合到 LLaMA2預訓練模型生成出具情緒浮動的故事，要注意的是 LLaMA2的輸出限制最多只能有 4096個token。我們將此產生過程稱為 MTSPL (Music To Story Procedure with LLaMA)。

Medication errors in patients are a global problem. They can negatively affect patients and be costly for hospitals and medical clinics. In 2021, a 28-year-old man with heart problems was admitted to a hospital in Porto Alegre. Due to a pharmacy error and insufficient monitoring in the administration, he received a dose 10 times higher than prescribed. This caused serious and probably irreversible damage to the patient. Reading the news and following the case in the media has encouraged research in scientific databases, searching for information and data on medication errors, as well as emerging technologies to reduce the occurrence of adverse medication events. Based on the findings of an English study that proved that errors occur at the drug prescription stage, the first stage of this research focused on drug dosage errors. The aim of this study is to develop an application based on artificial intelligence that can recognise these errors and help prevent them. The application uses a neural network to analyse prescriptions and warn of possible cases of incorrect dosage. The computer program was developed using a neural network and the drug dosage error recognition system using Python and Keras. The system was trained with 10 drugs and correct and incorrect dosage cases. A graphical interface was created to input and display new case data. Neural networks with different configurations were tested to obtain high accuracy with the training and validation data. A confusion matrix was used to assess the accuracy of the network for cases not used for training. The accuracy was approximately 96%, but problems were found in certain intervals. The errors are due to the need for more training, higher processing capacity and a cloud server. The results of the first stage of the research indicate the feasibility of using a neural network to recognise medication dosage errors and thus preventing the associated risks. Such a method could prevent cases like the one in Porto Alegre. Future studies could incorporate more types of drugs, allergies, drug interactions, pre-existing illnesses and other relevant factors into the system.

Over 30 million souls live in a world of darkness, a number greater than the populations of both Norway and Sweden combined. Every individual deserves the chance to embark on a journey across our magnificent blue planet. Yet, regrettably, little has been done to assist them. With this project, we’re lighting the way for the blind to explore our beautiful world independently, breaking free from dependence and embracing boundless horizons. In order to put our theory of the project into practice & explore the use of artificial intelligence & computer science, we started by collecting the required materials for our project such as micro-controllers, sensors, a pair of glasses, a laptop, and a miniature camera. Then we moved onto creating the project itself in which the digital software programmed onto the hardware plays the key-role, as the sensors and the camera will record the details and information from the surroundings and send it to the laptop for further processing. The camera would be the backbone of our project, as it will stream real-time footage to the laptop which will be analyzed by an open-source object detection model ‘YOLOv8’ for identifying objects. After finishing the base model of our project, we tested it in-front of objects such as toy cars, bikes, people, etc, and the results of the object-detection would be shown on the laptop. To observe this data, we created a device which has different modules and integrations for different functions. For example, we will use our camera and then stream it onto a laptop so the reading and the data can be processed on the laptop by AI using YOLOv8. As mentioned in the start, many people do not possess the ability to see, to assist them we have thought of this device which uses all readings and its analytical skills to analyze data and help them navigate, travel or simply, live a better life.

本研究探討 Instruction-Tuning 對法律領域語言模型的影響，我們使用 ChatGLM-2 6B 作為基礎模型,先以台灣法律文本進行 Continual Pre-training，再以和律師的 Q&A 數據集，分別採用 Supervised Fine-Tuning(SFT)、Reward Model 及 Proximal Policy Optimization(PPO)等 Instruction-Tuning 方法進行微調。結果顯示,僅經過 Pre-training 及 SFT 的模型,其產生的回覆較符合法律專業風格;但考量模型對法律知識的掌握,則以 Pre-training、SFT 及 PPO 整套 Instruction-Tuning 的結合效果最佳。本研究證明, 針對單一領域的語言模型, 不同的 Instruction-Tuning 方式會對其回覆風格及知識掌握造成不同影響。我們的研究為未來單一領域語言模型訓練提供了參考。

台灣少子化、老齡化問題迫切，醫療資源入不敷出，使得遠距醫療成為潛在的解決方案與趨勢。然而，遠距醫療的身份驗證安全性尚未完善。指靜脈辨識有非接觸、體內生物特徵等特性，其在衛生性和高安全性的優勢在醫學相關領域與醫療院所備受關注。若能夠開發出遠距醫療可用的指靜脈辨識，將有機會為遠距醫療產業的安全性貢獻一份心力。本研究主要分為兩階段：其一旨在優化指靜脈辨識技術，利用輕量化 CNN 指靜脈辨識模型，結合 Mini-RoI 技術， 使用 FV-USM 以及 PLUSVein-FV3 兩個資料集訓練我們開發的 FVeinLite 指靜脈辨識模型，並使用不同的 epoch 值訓練出最好的模型。而我們訓練的模型相較於其他指靜脈技術具有高辨識正確率、參數量更少、運算速度快等優勢。其二，我們將模型結合自製的低成本嵌入式裝置， 並製作 API 與並使用模擬的病患資料完整打造一個可使用於遠距醫療及醫療院所的指靜脈身分辨識系統。

"Blindness keeps you from things, deafness keeps you from people" (Helen Keller) Wibrazz is a communication tool that can be placed inside sportswear. Two versions have been developed. The simpler one allows hearing-impaired footballers to compete in the league with other athletes. The referee is given an additional device to give a signal when he blows his whistle. The hearing-impaired footballer then senses the signal from the device he is wearing and knows that he must pay attention to the referee. The complex version speeds up communication between the coach and the players during training sessions. It allows the coach to send simple messages to his players using his smart device. The athlete senses the signal from the device and acts on what has been previously discussed (e.g. a long signal means, "Everyone come to me!") With over 70 million deaf people worldwide, and 2-4 out of 1000 people in the United States who are functionally deaf, this can affect an individual's mental and physical well-being, and it is therefore a pressing issue to provide these athletes with the means to develop their talents in a traditional team environment. In addition to the organisations within countries, the ICSD is present on the international stage. Their importance is demonstrated by the fact that the 2023 Deaf Football World Cup featured teams from countries such as the United States, Germany, England and Japan.