本論文敘述對於智慧型個人數位相片管理瀏覽平台之研究、設計與實作過程。系統設計上基於整合多重證據架構，採用影像內容與使用者瀏覽行為之分析作為自動分類，判斷影像重要性與推薦程度的依據。影像自動分類方面，包括外部給予的標準資訊-EXIF資訊與分析影像內容，以其中人物存在數量與面積比例為依據的影像分類。而在影像的推薦方面，則採用影像品質之分析-包括對焦品質分析、曝光品質分析-與分析使用者瀏覽相片時的行為-包括停留時間與專注程度的整合為分析重要程度依據；最後則採用多模式(Multi-Modal)架構整合不同的評估結果並作為推薦的結論。 / In this thesis, we present the design and implementation of an intelligent personal digital photo browsing platform. The proposed system relies on multiple evidences inferred from image content as well as user behavior. Specifically, external EXIF data and face detection results are utilized to coarsely classify the digital images. Measures of image quality, including clarity and contrast, are calculated to further refine the search result. Moreover, we use web cameras to record and analyze the viewing behavior of the user and attempt to correlate the interest of the viewer to the effective viewing time. Finally, a multi-modal system is put in place to integrate the clues acquired from different modules.

誇張肖像畫是以誇張與諷刺的手法來表現模特兒的特徵。現今這類畫作比起以寫實手法繪製的肖像畫，更受到大眾的喜愛；但是人們若想要自己模仿繪製出這類的肖像畫，除了要有繪畫天賦之外，也必須經過長時間的繪畫訓練以及對人臉的觀察，才能抓出模特兒的臉部特徵，並用生動的手法誇飾出來。 因此，如果電腦能做到代替誇張肖像畫家的工作，自動模仿繪製出誇張化的卡通肖像畫，將能替人們節省大量的時間及經費。本研究主要在設計一套能誇張肖像畫的系統，以根據使用者輸入的臉部影像，自動擷取臉部的特徵（包括特徵點的相對位置、絕對位置、形狀及大小等），並藉由分析這些特徵與常人相異之處，以一致的方式自動將特徵誇張化。這個系統並能以藝術家的作品為輸入，將使用者的臉部影像轉換成具有畫家卡通風格的誇張肖像畫。除了人臉的效果外，我們也進行了頭髮分離與模擬的研究，以強化畫像風格模擬的完整性。最後，我們以實做出了系統對數個具特徵的人臉影像進行實驗，以驗證其可行性及有效性。 / Caricaturists are good at drawing sketches which express the exaggerated likeness of a person with a bit flavor of humor or sarcasm. People are willing to pay for this kind of work because it requires a lot of practice to achieve excellence. Acute observation is needed to extract the distinct features from the subject, and decent painting skill is essential to depict these features vividly. It will save much time and effort if computers can be trained used to draw caricatures. In this thesis, we developed a system which can extract and analyze facial features from simple an input facial images. The main purpose of this system is to generate the user-own caricature model by exaggerating his/her unique facial features. Different types of features, including relative locations and sizes, absolute locations and sizes, and each the shape of features are all taken in accountwill be considered. Unlike the complex process reported in the literature, we develop a transformative process that can handle different types of features in a more uniform fashion. Using an artist’s finished work as the source image; the proposed system is capable of producing cartoon-like colorful caricatures of a similar style effectively and efficiently. Besides the caricature painting of the face part, we also present some approaches for hair segmentation and hair style painting to increase make the system more the completeness of our system. Finally, we prove the feasibility and effectiveness of our system by showing several experimental results.

在不利的環境下做交通標誌的偵測與辨識是一件非常艱困的工作，無論在郊區或市區，複雜的環境、天候、陰影以及任何和光線有關的因素甚至是交通標誌遭到遮蔽都將使得偵測與辨識交通標誌變得相當困難。在本篇論文中，我們定義出較寬鬆的顏色分類(color thresholding)方法，配合一些交通標誌的特徵(如外形)來實作出召回率(Recall)較高的偵測系統，另外在辨識方面，最重要的是找出好的辨識特徵，因此我們利用離散餘弦轉換(discrete cosine transform)和奇異值分解(singular value decomposition)處理待辨識標誌擷取其特徵，並配合一些其他的交通標誌特徵，當作類神經網路(ANN)、naïve Bayes classifier等辨識方法的輸入，來幫助我們完成辨識的工作。目前實作出來的系統在有挑戰性的測試資料下有七成六左右的辨識率。 / Robust traffic sign recognition can be a difficult task if we aim at detecting and recognizing traffic signs in images captured under unfavorable environments. Complex background, weather, shadow, and other illumination-related problems may make it difficult to detect and recognize signs in the rural as well as the urban areas. In this thesis, I define a formula for color classification and apply other related features such as the shape of the traffic signs to implement the detection component that offers high recall rate. In traffic sign recognition, the most important thing is to get the effective features. I use discrete cosine transform and singular value decomposition to collect the invariant features of traffic signs that will not be severely interfered by disturbing environments. These invariant features can be used as the input to artificial neural networks or naïve Bayes models to achieve the recognition task. This system yields satisfactory performance about 76% recognition rate when I test them with very challenging data.

本研究嘗試不全以負面假設來看待配戴眼鏡對人臉辨識的影響。吾人將以邊緣偵測圖為基礎，以邊點強度的分析來建立一套定位眼鏡的偵測系統。同時用偵測出的鏡框位置，以邊緣點的強度、密度比較的方式，定位眼睛的位置；並以前述兩套偵測演算，採擷其過程的資訊，進一步地定位鼻子與嘴巴的位置。這些演算形成一個簡易的人臉特徵定位系統，其將可處理配戴眼鏡的人臉；吾人也將進一步地經由其處理過程與結果，分析眼鏡對區域人臉辨識的影響，進而引導出非自然物件可能對人臉辨識的阻礙或輔助。 論文也將以全域比對法中的PCA與ICA演算法作一連串的實驗，剖析眼鏡對於全域辨識的影響；此外，亦用相同的方法來測試非自然物(眼鏡)、光源亮度與人臉角度對於人臉辨識阻礙的程度，以探究是否系統值得花費更大的代價，來移除眼鏡這個被一致認定的人臉辨識障礙，並得以在辨識演算法上獲得更高的效能。 / The objective of this thesis is to investigate the efficacy of face recognition systems when the subjects are wearing glasses. We do not presume that non-facial features such as glasses are nuisances. Instead, we will study whether the inclusion of glasses will have a positive impact on the face detection procedure and how it affects the feature extraction process. We will demonstrate how to use techniques based on local feature analysis to reduce the uncertainties in the matching result due to interferences around the eyes and nose caused by optical glasses. We have also conducted extensive experiments to analyze the effect of glasses on face recognition systems based on global matching strategy. Specifically, we perform both principal component analysis (PCA) and independent component analysis (ICA) on face databases with different percentage of subjects wearing eye glasses. It is concluded that external objects such as glasses will have a negative impact on face recognition using global analysis approaches. However, the adverse influences of illumination and pose are more conspicuous during the recognition process. Therefore, one should take caution when attempting to adapt the global matching scheme to handle the difficulties caused by glasses.

自動化三維建模需要準確的三維點座標，而三維點的位置則依賴高精度的對應點，因此對應點的尋找一直是此領域的研究議題，而使用稀疏光束調整法(SBA：Sparse Bundle Adjustment)來優化相機參數也是常用的作法，然而若三維點當中有少數幾個誤差較大的點，則稀疏光束調整法會受到很大的影響。我們採用多視角影像做依據，找出對應點座標及幾何關係，在改善對應點位置的步驟中，我們藉由位移三維點法向量來取得各種不同位置的三維補綴面(3D patch)，並根據投影到影像上之補綴面的正規化相關匹配法(NCC：Normalized Cross Correlation)來改善對應點位置。利用這些改善過的點資訊，我們使用稀疏光束調整法來針對相機校正做進一步的優化，為了避免誤差較大的三維點影響到稀疏光束調整法的結果，我們使用穩健的計算方法來過濾這些三維點，藉由此方法來減少再投影誤差(reprojection error)，最後產生較精準的相機參數，使用此參數我們可以自動化建出外型架構較接近真實物體的模型。 / Automated 3D modeling of the need for accurate 3D points, and location of the 3D points depends on the accuracy of corresponding points, so the search for corresponding points in this area has been a research topic, and the use of SBA(Sparse Bundle Adjustment) to optimize the camera parameters is also a common practice, however, if there are a few more error 3D points, the SBA will be greatly affected. In this paper, we establish the corresponding points and their geometry relationship from multi-view images. And the 3D patches are used to refine point positions. We translate the normal to get many patches, and project them into visible images. The NCC(Normalized Cross Correlation) values between patches in reference image and patches in visible image are used to estimate the best correspondence points. And they are used to get better camera parameters by SBA(sparse bundle adjustment). Furthermore, it is because that it usually exist outliers in the data observed, and they will influence the results by using SBA. So, we use our robust estimation method to resist the outliers. In our experiment, SBA is used to filter some outliers to reduce the reprojection error. After getting more precise camera parameters, we use them to reconstruct the 3D model more realistic.

基礎矩陣在影像處理是非常重要的參數，舉凡不同影像間對應點之計算、座標系統轉換、乃至重建物體三維模型等問題，都有賴於基礎矩陣之精確與否。本論文中，我們提出一個機制，透過粒子群最佳化的觀念來求取基礎矩陣，我們的方法不但能提高基礎矩陣的精確度，同時能降低計算成本。 我們從多視角影像出發，以SIFT取得大量對應點資料後，從中選取8點進行粒子群最佳化。取樣時，我們透過分群與隨機挑選以避免選取共平面之點。然後利用最小平方中值表來估算初始評估值，並遵循粒子群最佳化演算法，以最小疊代次數為收斂準則，計算出最佳之基礎矩陣。 實作中我們以不同的物體模型為標的，以粒子群最佳化與最小平方中值法兩者結果比較。實驗結果顯示，疊代次數相同的實驗，粒子群最佳化演算法估測基礎矩陣所需的時間，約為最小平方中值法來估測所需時間的八分之一，同時粒子群最佳化演算法估測出來的基礎矩陣之平均誤差值也優於最小平方中值法所估測出來的結果。 / Fundamental matrix is a very important parameter in image processing. In corresponding point determination, coordinate system conversion, as well as three-dimensional model reconstruction, etc., fundamental matrix always plays an important role. Hence, obtaining an accurate fundamental matrix becomes one of the most important issues in image processing. In this paper, we present a mechanism that uses the concept of Particle Swarm Optimization (PSO) to find fundamental matrix. Our approach not only can improve the accuracy of the fundamental matrix but also can reduce computation costs. After using Scale-Invariant Feature Transform (SIFT) to get a large number of corresponding points from the multi-view images, we choose a set of eight corresponding points, based on the image resolutions, grouping principles, together with random sampling, as our initial starting points for PSO. Least Median of Squares (LMedS) is used in estimating the initial fitness value as well as the minimal number of iterations in PSO. The fundamental matrix can then be computed using the PSO algorithm. We use different objects to illustrate our mechanism and compare the results obtained by using PSO and using LMedS. The experimental results show that, if we use the same number of iterations in the experiments, the fundamental matrix computed by the PSO method have better estimated average error than that computed by the LMedS method. Also, the PSO method takes about one-eighth of the time required for the LMedS method in these computations.

奇異值分解(singular valve decomposition)是一個重要且被廣為運用的矩陣分解方法，其具備許多良好性質，包括低階近似理論(low rank approximation)。在現今大數據(big data)的年代，人們接收到的資訊數量龐大且形式多元。相較於文字型態的資料，影像資料可以提供更多的資訊，因此影像資料扮演舉足輕重的角色。影像資料的儲存比文字資料更為複雜，若能運用影像壓縮的技術，減少影像資料中較不重要的資訊，降低影像的儲存空間，便能大幅提升影像處理工作的效率。另一方面，有時影像在被存取的過程中遭到雜訊汙染，產生模糊影像，此模糊的影像被稱為退化影像(image degradation)。近年來奇異值分解常被用於解決影像處理問題，對於影像資料也有充分的解釋能力。本文考慮將奇異值分解應用在影像壓縮與去除雜訊上，以奇異值累積比重作為選取奇異值的準則，並透過模擬實驗來評估此方法的效果。 / Singular value decomposition (SVD) is a robust and reliable matrix decomposition method. It has many attractive properties, such as the low rank approximation. In the era of big data, numerous data are generated rapidly. Offering attractive visual effect and important information, image becomes a common and useful type of data. Recently, SVD has been utilized in several image process and analysis problems. This research focuses on the problems of image compression and image denoise for restoration. We propose to apply the SVD method to capture the main signal image subspace for an efficient image compression, and to screen out the noise image subspace for image restoration. Simulations are conducted to investigate the proposed method. We find that the SVD method has satisfactory results for image compression. However, in image denoising, the performance of the SVD method varies depending on the original image, the noise added and the threshold used.