- 學位論文
應用空間資訊於自動化彩色影像切割技術之研究
A Study of Automatic Color Image Segmentation Technology Based on Spatial Information
摘要
一張彩色影像通常包含許多不同的物件，因此彩色影像切割的主要目的在於將彩色影像明確的切割為數個具有相似特性或者可視為一體的物件或群組，透過影像切割的技術得到使用者有興趣的主體，當作物件辨識和影像查詢等前置作業。目前大多數的影像切割技術都需要人工設定參數值和辨識，屬於半自動化的設定。因此本論文中提出了兩種彩色影像切割技術，皆使用人眼最適應的色彩空間(RGB color space)並且能事先有效的自動取得不同影像所需設定的初始分群數，進而分別使用分群的切割技術得到相似於人工切割影像之最佳結果。 本論文第一個切割技術提出了一個簡單且快速的彩色影像切割技術，將輸入的彩色影像各自分離出RGB三層，在各自色頻下計算其變化量形成histon，再分析影像histon的累積量。Image histon是藉由觀察每個像素值與周圍鄰居的關係而形成新的影像累積圖。根據histon的變化量決定各色頻的切割點，並且計算其整張影像之均方誤差(Mean Square Error)與前一次變化量比較判斷是否收斂，之後再根據各頻帶的切割點組成不同的排列組合，重新將三個色頻(r,g,b)視為同一累積點，得到RGB histon，再計算RGB histon其變化量找出最終門檻值，並且有效的將彩色影像的物件明顯切割出來。 本論文第二個切割技術提出了一個有效的彩色影像切割方法，使用color histon的資訊應用於修改後模糊平均數群聚演算法(Improved Fuzzy C-Means, IFCM)演算法。Histon的建立是藉由同時取得RGB三個色頻的色彩資訊與周圍鄰居的空間資訊。首先將三個色彩空間頻帶的值各自量化並且組合成bins，此動作是為了得到影像中的參考色彩。接著為了得到空間資訊，計算中心點與周圍鄰居的相似度來形成histon。最後藉由計算histon中第i個 bin 和第i+1個 bin的距離來設定不同影像的初始分群數，並且運用於使用空間資訊的修改後模糊平均數群聚演算法來進行影像切割。
並列摘要
Since a color image usually contains a number of various objects, the main purpose of the color image segmentation is to clearly divide a color image into several objects or clusters where each object contains pixels that are similar or deemed to belong to the same group. Through the segmentation technique, one can extract objects that are interested by users or preprocess images for other operations such as object recognition and image retrieval. In this thesis, the first technique presents a simple and fast method to segment the color images. The method begins by extracting individual R, G and B color channels, and then the method proceeds to form histons by analyzing the frequency changes in each color channels. An image histon is the histogram formed by calculating the relationships between individual pixel value and that pixel’s neighboring pixels. With the frequency changes of histon in each channel, the method determines the thresholds and uses these values as segment points in each channel. Then, the method permutes segmenting points to generate all possible combinations of three segmenting points, where each point is selected from R, G and B color channel respectively. Every single set of R, G and B values will be treated as a new frequency point. These new frequency points form a new RGB histon. The final threshold values can be found by analyzing the frequency changes in this new RGB histon. The proposed method can segment objects in color images effectively. The second technique in this thesis presents an efficient algorithm to obtain the segmented images for color image segmentation. The proposed method employs the information of color histon and applies to the improved fuzzy c-means algorithm. A histon is compiled based on color information in RGB color space and spatial information among the central pixel and its neighbors. The proposed method first obtains new bins by quantizing each of RGB channels. The purpose of this procedure is to find representative colors of the image. The algorithm proceeds to calculate the similarity between the central pixel and its neighborhoods and build a histon in order to obtain the spatial information. As bins and the histon have been obtained, calculate the difference of the ith bin and i+1st bin in histon to set the initial number of clusters for the image. Finally, apply the processed spatial information to the improved fuzzy c-means algorithm for image segmentation.