電腦科學與資訊工程

本研究基於雙鏡頭立體成像原理，提出適用於長距離的相機光軸校正演算法則。此演算法則可快速地將失真的影像修復，以提供正確物體影像的視差資訊來得到其3D立體影像的資訊。由於物體影像於長距離時會太小而變得難以辨識，因此加大影像解析度能有效提升物體影像的辨識度進而獲得長距離的立體影像，但電腦執行影像處理的運算時間也隨之變長而失去即時性。因此本研究採用圖形處理器及多重執行緒的技術，可將解析度1280*720的影像擷取和處理的運算時間縮短至每幅影像100毫秒以下，而達到影像處理的即時性。本研究最後將此研究成果用於汽車防撞警示雷達系統設計，此系統運用兩個網路攝影機來擷取影像，並利用OpenCV所提供之開放原始碼實現本研究所有使用之影像處理程序，以快速達到汽車防撞警示雷達系統的實現。

Over the past year Conor has been developing an electronic time keeping device named Hourglass. Hourglass has a three-fold focus on functionality, intuitive design and simplicity. To simplify the device he has limited the hardware to a bare minimum. Just three buttons and an LCD screen comprise the user-interface. Although this interface is simple, the user can access many features. These include intuitive scrolling menus, countdown, lap and alarm functions, accessed through button combinations as well as multiple ways to use single buttons, such as holding or short pressing. Many functions have been integrated into the device, such as a stopwatch with lap times, a countdown, up to 99 Custom Alarms with an individual active/inactive state and a lock/unlock feature. The stopwatch is accurate to 1 second and can be started, stopped, reset and used to record lap times. When laps have been recorded, the user can then take the time value of a lap and turn it into a countdown. A countdown of up to 99 hours can be set, and will run until deactivated or until it reaches zero. Upon reaching zero the alarm is activated. The home screen displays the time, any active countdown and notifies the user if an alarm is active. It can be locked or unlocked by holding the blue button a set period of time, helping to reduce any inadvertent change in setting. All of the functions available can be operated easily with the intuitive 3 button interface method. The menu system is simple, but has been set up through clever coding. An arrow indicated which option is selected, by pressing the top button on the clock the option above the current selection is selected/the menu scrolls up. Pressing the bottom button selects the next option in the downward direction/scrolls down. The button in the centre positioned off to the left is used to activate an option. When a Yes or No prompt appears on the screen, the action corresponds with the button position. Therefore the triangle layout of the buttons is simple and intuitive. Thus Conor’s device relies on complicated, yet elegantly formulated and annotated code and simple hardware interfaces to interact with the user in a way which is intuitive and provides great functionality. It does this while being simple and easy to understand. Here these principles are applied to a clock project, but there are implications for good design that go way beyond this context.

File compression has become a very important tool in the technology field because it allows faster data transfer rates over the internet and decreased file size on data disks. File compression aims to reduce file size while still retaining the quality of the file. Lossy file compression methods are not very efficient because the compressed files end up losing more data than what is usually intended causing a considerable loss in quality. Lossless file compression methods, on the other hand, take time to process since they require decompression to retrieve the original file. In this study, a lossless algorithm which does no require decompression was created. The resulting Fractal File Compression (FFC) algorithm contains two parts, the IFS algorithm and the Huffman Tree generator. Both algorithms were created using Java language and JCreator. The finished program was tested on an image file with 2542 x 1944 pixels dimensions. The image file was compressed using JPEG, BMP, PNG and FFC formats. For each method, the image file was compressed at three different resolution settings; low, medium and high. All the compressed images were then viewed under 500% zoom using Adobe Photoshop CS2. In an area of 40 by 40 pixels, the number of distinct boxes, which served as a measurement of image quality, was determined. Compressed images for JPEG, BMP, and PNG for both the low and medium settings have low image qualities, while the fractally-compressed images have a high image quality. For the high resolution setting, both JPG and fractally-compressed images have Page 2 of 2 high qualities while BMP and PNG still have low qualities. Based on the measurements obtained from the box-counting method and the file sizes, the absolute image quality for each compressed image was calculated. The absolute image qualities of the compressed images used for each setting were then compared. Coupled with large file size and small pixels per area count, the conventional methods have lower absolute image quality than the images compressed using the FFC method. This was true for the low and medium settings, however, JPG compression has a higher absolute image quality than the fractally-compressed images. This meant that JPG compression is more efficient than fractal compression when an image has a high resolution. The resulting FFC algorithm is lossless since it uses pattern searches and replacements in order to decrease the file size. To make the program more suitable for high resolution images, the FFC algorithm may be modified. Most of the changes in the FFC algorithm should be done in the IFS generator. High resolution images can be compressed fully if the pattern that was used for compression is more representative, but even shorter. A more representative bit pattern would create a high quality, high resolution image with a smaller file size.

Despite the multitude of biomedical scans conducted, there is still relatively low accuracy and standardization of diagnoses from these images. In both the fields of computer science and medicine there is very strong interest in developing personalized treatment policies for patients who have variable responses to treatments. The aim of my research was automatic segmentation of brain MRI scans to better analyze patients with tumors, multiple sclerosis, ALS, or Alzheimer’s. In particular, I aim to use this information, along with novel artificial intelligence algorithms, to find an optimal personalized treatment policy which is a non-deterministic function of the patient specific covariate data that maximizes the expected survival time or clinical outcome. The result of the research was IlluminaMed, a biomedical image analysis toolkit that relies on the development of new artificial neural networks and training algorithms and novel research in fuzzy logic. The networks can detect patterns more complex than humans can identify and create patterns over long periods of time. IlluminaMed was trained by a dataset of professionally and manually segmented MRI scans from several prestigious hospitals and universities. I then developed an algorithmic framework to solve multistage decision problem with a varying number of stages that are subject to censoring in which the “rewards” are expected survival times. In specific, I developed a novel Q-learning algorithm that dynamically adjusts for these parameters. Furthermore, I found finite upper bounds on the generalized error of the treatment paths constructed by this algorithm. I have also shown that when the optimal Q-function is an element of the approximation space, the anticipated survival times for the treatment regime constructed by the algorithm will converge to the optimal treatment path. I demonstrated the performance of the proposed algorithmic framework via simulation studies and through the analysis of chronic depression data and a hypothetical clinical trial. IlluminaMed can automatically segment the scans with 98% accuracy, find tumors with 96% accuracy and approximate their volume within a 2% margin of error. It can also find lesions in MS and ALS, distinguishing them from tumors with 94% accuracy. IlluminaMed can, in addition, determine the tendency of a patient to develop Alzheimer’s several months before patients develop symptoms correlating the brain structure and its fluctuations. Lastly, the censored Q-learning algorithm I developed is more effective than the state of the art clinical decision support systems and is able to operate in environments when many covariate parameters may be unobtainable or censored. IlluminaMed is the only fully automatic biomedical image analysis toolkit and personalized medicine engine. The personalized medicine engine runs at a level that is comparable to the best physicians. It is less computationally complex than similar software and is unique in the fact that it can find new patterns in the brain with possible future diagnoses. IlluminaMed’s implications are not only great in terms of the biomedical field, but also in the field of artificial intelligence with new findings in neural networks and the relationships of fuzzy extensional subsets.

西爾平斯基船帆(Sierpinski Gasket)與西爾平斯基掛毯(Sierpinski Carpet)都屬於碎形(fractals)圖形的一種，可以利用迭代運算系統IFS(Iterated Function Systems)碼來產生，代入迭代運算方程式後，經由多次的運算，可以得到重覆的圖形。本研究中，我將提出一些作法，找出西爾平斯基船帆與掛毯圖形其遞迴關係式，進而討論出其數位圖形之規律性及所涵蓋的內容與性質，著重在推廣西爾平斯基船帆與掛毯圖形的概念，將一段音樂曲取出，把它們看成反覆隨機迭代點，利用程式經由多次的插值運算，計算出各段音符。最後加入基因演算法來解決音符長短的問題，把製造好的音符染色體放置到交配池中，以隨機的方式在交配池中選取其中之一個染色體進行交配的動作，此二音符染色體會交換彼此的基因，產生下一代新的代表音符長短之染色體，應用於數位音樂創作，而衍生的西爾平斯基船帆與掛毯圖形新穎應用與創新的結果，希望能提供數位音樂創作的多樣性，更進而可以找出「好聽的音樂」與數學的直接關聯性。‘Sierpinski Gasket’ and ‘Sierpinski Carpet’ are two graphics that belong to fractals. They can be produced by IFS (Iterated Function Systems). By iterative computation of many times, we can obtain the similar graphics. In my research, there are some methods to generate Sierpinski Gasket, Sierpinski Carpet, and the iterative algorithms. In addition, I would discuss the regularity and the content as well as the properties of those digital patterns. At last, the advanced application of Sierpinski Gasket and Sierpinski Carpet to digital music pieces was presented. The program took a note of several measure of music as the beginning point, and made the IFS calculations for each new note in each measure. But there was no difference in beats if you just make the IFS iteration. So I changed the beats with genetic crossover method. In this research, the expression of the DNA to each beat of note was adopted. The same way, it took a note as a beginning point. And the system obtained the new DNA from the old notes for new ones randomly. That would make a piece of brand new music. What I want to do in this research is improve the multiformity of music and find what the relationship is of ‘good music’ and mathematical algorithms.

Our goal is to make roll call systems at schools technological. While Rolling Call system is getting technological capacities some useful outcomes occur as well like; To remove the cost for class books by abolishing the class books used for roll call To eliminate mistakes with the usage of class book .(The numbers written incorrectly) To save again a large amount of waste paper( The short messages will be send to the parents instead of mailing the letters to the their adresses once a week about their students' absance.) To improve communication among school and parents ,by this way to prevent various kinds of problems arising from absenteeism. To make teachers school boards more efficently and motivated besides their affairs will get easier as well. To save time of school managers(They have no Works like saving the absenteeisms into e-school network because everything including all the absenteeism info will be carried into the system automatically) To save time of each lesson the teachers will not loose time for the roll call ,so they will be able to devote their time to their students and training. -Technological polling mechanism will contribute to education, this contribution will make all the students and teachers happy because all the teachers carry concern whether their curriculum will be finished or not through this system recording all the rests of teachers because of illnesses and some national celebrations.

本研究主要利用最大概似法及動態規劃演算法來嘗試縮短生醫領域在疫苗研發的時程。透過序列比較的計算方式加速找出病毒序列具有專一性的有效區段。使科學家可以減少盲目測試的實驗。我們期望找出經過電泳之後，可以判斷具有可製造疫苗的最佳生物序列區段。藉由已知流感病毒的基因序列來分析現有流感病毒的演化親緣關係。嘗試由已知流感病毒疫苗來設計未知的流感病毒疫苗之建議(結果如下圖)。

在數位化的今日，由於網際網路的技術蓬勃發展，網際網路變得更容易使用及具高度的親和性，使得網際網路的使用逐年成長。隨著越來越多人依賴網路進行交易，也衍生了層出不窮的網路詐騙問題。其中，網路釣魚就是一項著名的詐騙技術：詐騙者透過偽裝成知名企業的網站，藉此騙取使用者的個人私密資料。在本研究中，我們提出了一套植基於彩色視覺密碼學原理的網站驗證機制，使用者可以透過此機制，直接利用人類視覺的方式來驗證所連上的網站是否有問題，並在此機制之下，設計出另一套管理使用者密碼的方式，進而方便使用者不必費心的去記憶密碼。 Recently, as networks technology flourishes, Internet becomes easier and friendlier to use, and makes the usage of Internet grow up year after year. With more and more people relying on online transactions, it leads to endless network fraud issues. Among them, phishing is a well-known fraud technology to disguise the famous business website to get user’s private information by cheating. Therefore, in this study, an effective scheme based on color visual cryptography is proposed to test and verify the website. Through the proposed mechanism, users can check whether there is a problem website by using human vision directly. Furthermore, the proposed scheme also provides another way to manage user’s password effectively.

Pediatric cancer has an incidence rate of more than 175,000 per year with a mortality rate of approximately 96,000 per year. One major cause of this problem is late diagnosis. A novel promising way of pediatric cancer screening is handwriting analysis. This method surpasses other methods by detecting pediatric cancer in a very early stage. However, studies are still limited to manual analysis which needs an expert and a long period of time. The aim of this project is to design a computer program to extract handwriting features and build a classification model to classify the user as patient or as control. Dataset was collected from schools and hospitals where all participants could read and write in English. After data cleansing, number of samples was 440 samples. MATLAB (Matrix Laboratory) program was used for extracting geometric features in handwriting. Program was validated using a subset of 50 samples of the dataset. WEKA Package was used to test and build the classifier. Experiments were done using classifiers: Logistic, Multilayer Perceptron, J48, LibSVM, AdaBoostM1 and Naïve Bayes. Best subset of attributes was evaluated and used for each classifier and all calculations were done as the average of cross validation operations of several folds assignments. Best performance was achieved by Logistic classifier with average accuracy of 80.15%, standard deviation of 0.43% and Matthews's correlation coefficient of 0.59. Finally, this project presents a new fast, free, ready, easy and psychologically comfortable method for pediatric cancer detection while keeping suitable accuracy for mass screening.

快門，捕捉最原始的感動；科技，創造最完美的呈現。當攝影遇上科學，成就了本研究的主題──高階電腦數位影像之研發。在我的研究過程中困難重重，從外景攝影、後製研究、影像創作到本研究撰寫完工，一切由我個人獨力進行，經歷了多次失敗，仍堅持的完成這重大的研究突破。 相信多數人會使用相機，但是對數位攝影這領域卻不甚了解，更別談藝術與科學的結合。因此希望本科展研究將會成為未來科技數位電腦的主軸，強調科技、即時、便利、環保、生活及教育的科學推廣，以實用並超越新世紀的數位領域，讓所有電腦愛好者，都可以輕鬆的應用此高階的數位暗房後製。讓電腦科技不只是零與一的組合，而是心靈與世界的互動！ ;Camera shutter captures the most original affection; technology creates the most perfect display. High-level digital image—when photography meets technology—is the theme of this research. However, in the process of the research, I have encountered many problems—from outdoor scene shot, production research, image creation to report writing—I did them individually and had undergone many failures, but I still insisted on finishing this big research breakthrough. I believe that many people can use the camera, but didn’t understand anything about the field of digital photography, let alone the combination of art and technology. I hope this technology development research would become the main perch of digital technology in the future—emphasizing technology ,instantaneous,convenience,environmentalprotection, livelihood, and popularize technology education—to use and go beyond the new age of digital field and to make computer lovers apply high-level darkroom production easily. Making computer technology not just combinations of zeros and ones, but the interaction between life and the world.