以統計方法為基礎之數位影像擷取技術

Abstract

為因應大量圖片的累積所伴隨而來的管理需求，在本論文裡我們提出了數種數位圖庫及個人數位相簿的管理技術。首先，對於數位圖庫的管理，我們提出了一個新的擷取色彩特徵的方法，此方法可保有原圖片的色彩分佈。此方法包含有FC和VC，是以BQMP為基礎的擷取方法，可保有圖片的色彩分佈並減少擷取過理中所造成的失真。除了以EMD作為此色彩特徵的距離函式，我們提出了一個快速且有效的距離函式，稱為CHIC，我們也探討了如何有效率地實作此距離函式。最後我們以內容為基礎的圖片搜取應用來評估此色彩特徵擷取方法的實用性。 此外，在本論文裡，我們也對個人數位相簿管理的問題深感興趣。我們首先探討了個人行動裝置上線上事件偵測的問題，有別於前人的研究所作之桌上型電腦的假設，我們著眼於在個人行動裝置上支援此應用。我們提出了一個同時考慮時間和地點的事件偵測方法。我們以Poisson隨機函式並加以適當的限制來模式化使用者的照相行為，當以此函式所作之偵測決定的可信度不夠高時，地點資訊會被一並納入考慮。 之後，我們探討了如何找出一組代表性的圖片。此問題為接續著事件偵測之後的下一個階段問題，找出一組代表性的圖片可以幫助使用者了解個人相簿的組織結構。我們認為一組代表性的圖片必須是每一張圖片儘量具代表性同時相鄰之間的圖片極具差異性。針對提供個人行動裝置上的應用，我們將此問題模式化成一個最佳化的問題並以Viterbi演算法求其最佳解。我們也進一步將此問題一般化於提供桌上型電腦上應用的支援。 最後，我們討論了個人相簿的圖片分類的問題。我們提出一個新的架構以解決此問題，它利用PAM叢集方法以幫助使用者標示學習資料，並以改進式的分類樹方法分類圖片，因為對使用者而言分類樹結構是易於呈現而了解的，此外，為了讓使用者對分類樹的結構給予回饋，我們將每一個節點連結一個代表圖以呈現分類結果。

Motivated by the need of managing a large number of images, we propose in this dissertation several techniques to support the management of personal photo collections and general image repositories. First, we propose methods that extract color features by preserving the image color distribution for the management of general image repositories. Based on the Binary Quaternion-Moment-Preserving thresholding technique, the proposed extraction methods, Fixed Cardinality and Variable Cardinality, are able to extract color features by preserving the color distribution of images and to substantially reduce the distortion incurred in the extraction process. In addition to utilizing the Earth Mover's Distance as the distance measure of our color features, we also devise an efficient and effective distance measure, Comparing HIstograms by Clustering. Moreover, the efficient implementation of our extraction methods is explored. We then evaluate the meaningfulness of the new extraction methods by the application to content-based image retrieval. In addition, we are interested in the problem of managing personal photo collections. We first explore the problem of online event detection on a mobile platform as opposed to most prior works in which a desktop platform is assumed. We propose an event detection algorithm that fuses time and location information, which is deemed the important information for personal photo management. Specifically, we model event occurrences in a user's photo-taking behavior as a Poisson process by imposing certain constraints on calculating the elapsed time. Location information is incorporated into event detection when confidence in a decision based on the Poisson process is not high enough. We next explore the problem of selecting a set of representative photos. Selection of representative photos can be deemed as the next step after event detection because a good selection can help users indicate the organizational structure of a personal photo collection. We consider that a good selection is one in which the selected photos are as representative as possible of their individual group and maximally distinct from one another in the distinction window. Motivated by the need for photo management support on mobile platforms, we formulate the problem as an optimization problem and solve it by the Viterbi algorithm. We further generalize the problem and propose appropriate algorithms to support photo management on desktop computers. We also study the problem of image classification on personal photo collections. We propose a framework, User-assisted image Classification on Personal photo collections, to address this problem, where the PAM clustering method is used to help a user label the training data. A modified decision tree classifier is proposed to classify photos because a decision tree classifier is easily understood by a user. In addition, we visualize the classification results by associating a tree node with a representative image, which facilitates a user to give feedback to our decision tree classifier.