對於以內容為主的資訊擷取的需求愈來愈高時，以畫面為主的傳統方法以不適用。新的多媒體應用正朝向基於以物件為主的視訊，一個視訊序列指包含人們感興趣的前景部分而沒有背景的部分來支援更靈活的應用。 例如 MPEG-7已經定義出提供使用者依據物件形狀作視訊資料搜尋的標準，而在MPEG-4 中也訂定了內容性應用（content-based）的功能，它將連續的影片拆分成一到數個影像物件平面（video object planes），簡稱VOP’s ，每個VOP 代表各一個移動的物件，如此一來可以對他們加以重新組合成一部新的影片或者針對他們的形狀去作壓縮處理。由此可知，發展出能從一般影片擷取出物件的技術是非常重要的。 在這篇論文裡，我們提出以區域為主的視訊切割演算法。我們利用不同尺寸的型態學特徵擷取及高階統計來分割影像中的物件。不同尺寸的形態學特徵擷取考慮特徵的尺寸和對比。而高階統計則非常適用於偵測動量為小的物件，以判斷是否遵守高斯分布來擷取移動資訊。根據實驗結果，這種方法對不同種類視訊都能提供不錯的結果。 As the demand for content-based information retrieval goes high, traditional “frame”-based videos are not adequate. Novel multimedia applications are looking for object-based video, a video sequence has only one object without background, to support flexible utilization. For instance, MPEG-7 (Moving Picture Experts Group) has defined standardized functionality that allows users to search visual content according to object shapes. Meanwhile, MPEG-4 video standard verification model includes the content-based functionality to decompose a video sequence into one or several video object planes (VOP’s), so that each VOP represents one moving object, and they can be recomposed as a new video sequence or be compressed according to their shapes. Therefore, to develop the technique of extracting objects from plain videos is very important. In this thesis, we proposed a region-based segmentation algorithm. It is based on multiscale morphological feature extraction followed by a higher order statistical test ( HOS ). Multiscale morphological features extraction, which takes the feature size and contrast into account for region extraction. The HOS algorithm is suited for very small moving because of the characterization, that suppress the statistic of Gaussian-distributed and enlarge the statistic of Non-Gaussian-distributed.video. It provided reasonable VOP extract procedure without simplification step and suit for very small moving objects extraction. Experimentally, this method provides good results on different kinds of sequences.
