数据加载中.....
|
jsp.display-item.identifier=請使用永久網址來引用或連結此文件:
https://tkuir.lib.tku.edu.tw/dspace/handle/987654321/35681
|
| 题名: | 基於特徵連續性補填之前景物件切割技術 |
| 其它题名: | An efficient video object segmentation based on mask pre-filling algorithm |
| 作者: | 張智越;Chang, Chih-yueh |
| 贡献者: | 淡江大學電機工程學系碩士班
余繁;Ye, Fun |
| 关键词: | 物件切割;影像型態學;Edge detection;mathematical morphology;video object segmentation |
| 日期: | 2007 |
| 上传时间: | 2010-01-11 07:00:09 (UTC+8) |
| 摘要: | 隨著網路快速發展及儲存技術的進步，人們對於視訊影像的要求除了原有的品質外，也開始增加對於互動性的需求，所以在新一代的視訊標準MPEG-4中更是首先採用以物件為基礎的編碼方式，而視訊物件的切割是當中關鍵的技術，過去被提出的演算分成兩類，第一類主要是在時間軸上利用兩張連續圖框間之差異取得物件的移動資訊，此法雖然擁有較低的運算量，但由於是根據直覺的考慮其差異度強健性較弱，而且欲切割之影像容易受外界環境所影響，如亮度與陰影等，切割品質較差；另一類方法結合時間軸與空間軸同時進行分析處理，也是較常見的方式，在此類方法中有常見有兩種作法，第一種作法，通常利用邊緣偵測方法的演算法找出圖框中物件的輪廓，之後再利用時間軸上的分析找出移動區域的邊緣，最後再將前景物件執行填滿的處理，形成物件遮罩。雖然此法雖有強健性較強和切割品質較佳之優點，但其在遮罩補滿處理的過程中會利用型態影像學運算來達成，往往會因預先選用的結構元素不同而影響遮罩與原物件近似程度，不僅如此，一旦選用較大或較複雜的結構元素時會增加處理的運算量導致系統處理速度的效能之降低；另一種方法是利用分水嶺法或色塊分割法，先將畫面切割成多個相似區塊再進行合併及移動偵測找出前景物件，不過由於此類方法會有著過度分割的問題，會造成在之後的處裡過程中物件形狀與特徵的辨視困難度，而且如何將過度分割的結果有效的將歸屬同一物件的區塊合併也是相當大的考驗。
因此，綜合上述的各種方法的優缺點考量下，本文所提出之切割方法，改善了前面所提及利用邊緣偵測配合移動資訊的方法，根據物體的連續性在以往傳統方法中將移動物件輪廓填滿的步驟前加入一個前處理，用以加強移動區域的輪廓邊緣，來符合效率的需求，並獲得較佳的切割效能。
As the technique of storage media and Internet develops rapidly in recent years, the demands for video are not only quality but also interaction. The conventional video coding standard, such as MPEG1, MPEG-2, and H.263 cannot satisfy the demands mentioned above .The MPEG-4 video coding standard is the first one to support randomly accessing video objects by the concept of video-object-plane (VOP). It can support high interaction and more flexible coding. Therefore, segmenting the shapes of the video objects is very important. Many video object segmentation algorithms have been presented. They can be summarily classified into two types: (1) Temporal Analysis and (2) Temporal-Spatial Analysis. One typical kind of Temporal Analysis is to get the moving information of video objects between two continuous frames on the temporal domain. Although these methods have lower computation load, the quality of segmentation is not good enough. Because the objects needed to be segmented are easily to be influenced by brightness or shadow, some kinds of typical method combine spatial and temporal domain was proposed [14]. This approach uses the algorithms of edge detection to get the shapes of the foreground objects and find out the edge of moving regions by using the analysis on the temporal axis, and then, the filling technique is used to generate the masks of the foreground objects. Although this method has higher robustness and better quality for segmentation, the similarity between the original video objects and masks depends on the chosen morphological structuring element and times of processing during the process of filling masks by using morphological operations. If we choose a complex structuring element, it will raise the computation load and reduce the efficiency of the system. In order to overcome the mentioned shortcomings, we propose a new algorithm by adding a pre-processing mechanism to improve the object segmentation.
A new video object segmentation algorithm using the morphological technique is proposed. Several video object segmentation algorithms used mathematical morphology to generate the object masks, but operations of mathematical morphology have two drawbacks: (1) high computation load, and (2) the quality of masks depends on the chosen morphological structuring element. There are many techniques to speed up the morphological operations by hardware implementation, but no discussions are found about reducing the influence of the choice of structuring elements. By adding a pre-processing mechanism, we effectively reduce the influences of the chosen structuring elements based on continuity of shape features and times of morphological operations. Experimental results indicate that our algorithm can improve the speed of filling operations of the object masks and accuracy of segmentation |
| 显示于类别: | [電機工程學系暨研究所] 學位論文
|
文件中的档案:
| 档案 |
大小 | 格式 | 浏览次数 |
|---|
| 0Kb | Unknown | 291 | 检视/开启 |
|
在機構典藏中所有的数据项都受到原著作权保护.
|
