使用空時域疊代法強化分析之視訊分割

Abstract

隨著科技的進步，人們對影像的需求日增月益，其應用範圍也相對的擴大。 如影像的壓縮、辦識與合成等。 為了滿足動態影像的相關應用， 一套理想的動態視訊區域分割技術是必需的。 對於動態影像區域分割的技術而言， 最重要的是要對區域或物體做一個明確的定義， 適當的定義將對區域分割的研究會有很大的助益。 本論文提出一套動態影像區域分割的方法。 首先介紹一個靜態影像分割的技術JSEG， 並運用之產生我們的初始單張影像分割。 接下來由貝氏動態估計與馬克夫隨機場域推導出 三個簡化的疊代步驟並對連續兩張頁框進行。 第一步進行動態影像中的動態向量的修正， 第二步的目地為找出將被摭住的區域， 經簡化後以一直觀的方化實作之。 第三步經簡化實行物件邊緣的修正與改進、 可細分為區域分併與邊緣修正。 經過對table tennis、flower garden等影像測試後， 雖然比就單張分割有所改進， 但其結果不如預期般理想。 將來可就增加參考張數與進一步實行貝式模型來改進。

With the advance of technology, our demand for video is increasing. As a result, the video application is expanded at the same time. Therefore, to employ the motion video to the full, we have to develop an ideal set of segmental technology as a preprocess. For the segmental technology, the most important is to precisely define the region or the object. A clear definition will be helpful to the research in the regional segmentation. In this thesis, we present a set of video segmentation methods. After introducing a still segmentation method JSEG, we use it to generate our initial single frame segmentation. Then we use Bayesian motion estimation and Markov Random Field to formulate our three iterative stpes between two continuous frames. The first step update the motion fields of video, and the second frame tend to find the occlusion region, we use a heuristic method to substitude the original step for simplification. And after simplification, the final step we modify the region fields, it can be divided into two parts: merging the regions and improving the accurate region boundaries. After simulation with the table tennis and flow garden sequences. Although there are some improvement compared to the original segmentation, the results do not as perfect as what we anticipate. In the future, we can take more frames into account or implement the Bayesian iteration without simplification.