以影像壓縮為導向的新式影像切割技術

影像切割是在影像處理領域中一項重要的課題，除了已廣泛被運用在圖像辨識外，也常用於影像壓縮前端的處理。在此論文中前面的章節，第二章到第三章中，我們先列出目前被廣泛使用的影像切割技巧，從影像切割最基本的概念切入，由區域概念的分割法介紹到以邊緣偵測概念的影像切割法，同時我們會簡單介紹一些已廣泛被應用在統計上並有可能拿來運用在影像切割的分群方法。第四章我們加上了對影像邊緣紀錄的研究結果，此章節可說明紀錄影像邊緣的重要性。在第五章的前段，我們選出三種極具特色且為目前廣受歡迎的影像切割法：種子區域成長法(Seeds Region Growing Algorithm)、K-means分群法(K-means Algorithm)、分水嶺演算法(Watershed Algorithm)，針對這三個影像切割方法做一個詳細的討論，並仔細的探討各方法的優缺點。在此論文中，我們將以影像壓縮前端處理的角色來定義影像切割。作為一個影像壓縮前端的影像處理角色，我們希望最佳的影像切割方法必須具有三項優點。第一項是速度優勢，第二項我們希望影像切割結果的各分區形狀必須具有完整性，第三項我們希望在前兩項優勢的前提下，仍能保有不錯的影像切割可信度。關於速度的取捨。由於影像切割的後端是概念較為複雜的影像壓縮端，我們希望將整個影像處理過程所花費的處理時間盡量保留給影像壓縮端，在此前提下，我們希望最好的影像切割法必須保有速度的優勢，意即，在影像切割端不需浪費太多處理時間。關於第二項，我們希望分割結果的各區形狀必須保有其完整性而非支離破碎，支離破碎且分散各處的分區形狀在我們送入後端影像壓縮的時候會有一些缺點，最大的缺點就是壓縮端必須額外花費資源來紀錄這些破碎區塊的邊緣，這會造成處理時間過長以及影像壓縮端的麻煩。區域成長法具有分割詳細，可信度高且分割形狀完整不破碎的優點，但他的缺點是處理時間非常慢。K-means分群法的優點在於其承襲了切割分群(Partitional Clustering)的優點：快速，但運用K-means做的分割結果，各區的形狀破碎且分散各地。分水嶺演算法的優點是快速，且分割詳細，但同時有過度分割的缺點。縱觀以上三種方法，都不能同時保有我們所需要的三種特性，因此我們在第五章的後半提出了一個新的方法，我們結合區域影像切割的特性與k-means分群法中重心轉移的優點來做影像切割，在新的方法中，我們擁有極快的切割速度，保有切割分區的完整性等優點。在第六章中，我們針對新方法進一步做了改良，我們將一張圖分為四乘四總共十六個分區，在十六個分區分別計算其變異數(variance)與平均頻率，以此兩種特性我們決定影像切割的閥值(threshold)。第七章是此論文的結論與未來展望，而在參考資料中，也對目前相關領域的研究做了一個分類整理。

關鍵字

影像處理；快速影像切割；邊緣偵測；傅立葉運算子

並列摘要

In this thesis, we define the role of image segmentation as the front-stage processing of the image compression. Base on this, we hope there is an image segment algorithm with three advantages which are the fast speed, the good shape connectivity of its segmenting result, and the good shape matching. In chapter 2, we introduce some region-based image segmenting algorithm, including hierarchical data clustering, partitional data clustering, region growing algorithm, and splitting and merging algorithm. In chapter 3, we introduce the edge-based image segmenting algorithm and take the watershed algorithm as an example. In chapter 4, we introduce the improvement of the Fourier descriptor. It helps us to realize the importance of recording the boundary of an image shape. In the front of chapter 5, we compare the three algorithms below, Seeds Region Growing, K-means algorithm, and Watershed algorithm. We compare them and discuss the advantages and the drawbacks of them base on to be the compression-oriented image segmenting algorithm. The region growing algorithm is very reliable in shape matching with the good connectivity of its segmenting result, but it is too slow for us to use it. The K-means algorithm is very fast, but its segmenting result is fragmentary. It is very inconvenient for us to record the boundaries of the results after we sent the results into the compression stage. The watershed algorithm is widely use in image segmenting, but it is also famous, and maybe, notorious for the over-segmentation result. In the end of chapter 5, we propose and introduce our image segmenting algorithm in details. Our algorithm owns the advantages of the fast speed, the good shape connectivity with its segmenting results, and the not bad shape matching. In chapter 6, moreover, we improve our algorithm by using the adaptive threshold selecting with the local variance and the frequency. It help us to segment the image with suitable threshold value and improve the shape matching of our image segmenting method. There are conclusions and future work in chapter 7. May this thesis be helpful for you.

並列關鍵字

image processing ； fast image segmentation ； edge detection ； fourier descriptor

參考文獻

A. Submitted Papers

[1] J. J. Ding, J. D. Huang, C. J. Kuo, W. F. Wang, “Asymmetric Fourier descriptor of non-closed boundary segment,” CVGIP, 2008.

[2] J. J. Ding, J. D. Huang, C. J. Kuo, W. F. Wang, “Asymmetric Fourier descriptor of non-closed boundary segment,” submitted to ICIP, 2009

B. Digital Image Processing

[3] R. C. Gonzalez and R.E. Woods, Digital Image Processing 2nd Edition, Prentice Hall, New Jersey, 2002.

國際替代計量

以影像壓縮為導向的新式影像切割技術

全文下載

主題瀏覽