Title: | 應用於彩色影像切割之區域性可靠資訊匯集 Local Belief Aggregation for Color Image Segmentation |
Authors: | 詹景竹 Jing-Chu, Chan 張添烜 Tian-Sheuan, Chang 電子研究所 |
Keywords: | 彩色影像切割處理;馬可夫隨機場理論;信任傳遞;Color image segmentation;MRF;Belief propagation |
Issue Date: | 2008 |
Abstract: | 在彩色影像切割裡,馬可夫隨機場理論被用來解決如何給予畫面像素適當標籤的問題。在此論文裡,我們以區域的內部特性以及區域和區域之間的相關性來建立起馬可夫模型。然而,龐大的切割標籤數量,對於使用信任傳遞 (Belief Propagation, BP) 演算法來近似以馬可夫隨機場理論為基礎的彩色影像切割法遇到一些困難。這些困難包含了以下兩點:計算複雜度過高以及記憶體儲存空間過大而不敷使用的問題。在此論文裡,我們另外提出了一個利用地域性可靠資訊匯集的演算法來解決這些問題。這個方法主要是以限制鄰近點傳送進來的訊息數量為概念來達成目的。我們將此演算法套用到我們提出來的馬可夫模型上,利用近似的方式找出最大事後機率 (maximum a posteriori, MAP) 的結果。跟原本的信任傳遞演算法比較起來,我們提出的演算法可以減少相當多的記憶體儲存空間。在評量影像切割的結果方面,我們選擇與眾所皆知的平均位移 (Mean Shift) 演算法來做比較。在此,我們使用非監督方式的評比方法。這個方法主要是利用色彩視覺差異的特性設計而成的。實驗數據顯示,所提出的彩色影像切割演算法無論在主觀或是客觀的評比上,皆可以得到與平均位移演算法有相似的效果。除此之外,所提出的演算法在運算方面也比平均位移演算法還來的更具平行性。 Markov Random Field (MRF) is used to solve the problem of labeling pixels in image segmentation. In this thesis, we formulate the MRF model based on the intra and inter region criteria. However, the enormous number of segment label in color image segmentation causes MRF-based color segmentation algorithm using belief propagation (BP) to suffer from complexity and storage explosion. To cope with this problem, this thesis also proposed a local belief aggregation (LBA) algorithm which restricts the number of messages to be aggregated from a neighboring node, to find the segmentation image that approximate the MAP solution of our MRF model. With the proposed LBA, memory storage is much reduced compared with the original BP algorithm. To evaluate the segmentation results, we compare the segmentation image with the well-known mean shift algorithm. Here, the unsupervised evaluation scheme using visible color difference is used as our objective evaluation metric. Experimental results show that the proposed color image segmentation algorithm can achieve a comparable result to mean shift algorithm both objectively and subjectively. Besides, the computation of LBA possesses more parallelism than the mean shift algorithm. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#GT009511638 http://hdl.handle.net/11536/38161 |
Appears in Collections: | Thesis |
Files in This Item:
If it is a zip file, please download the file and unzip it, then open index.html in a browser to view the full text content.