影像分割技術逐漸由只對一維灰階影像朝向三維空間彩色影像分割趨勢,影像分割技術的演算法中,群聚法在彩色影像具有優勢,因其考量三維度的資料;然而,一般利用群聚法的彩色影像分割演算法中,通常只考慮到色彩特徵的空間分佈,少數考慮到影像域內的空間特徵,而容易產生過度分割的情況發生。針對此點,本研究提出兩段啟發式彩色影像分割演算法;前半段,先應用能大量縮減資料量的類神經工具SOM,對色彩空間特徵進行分割;接續前半段之結果,在後半段中,利用影像區塊特徵進行合併的處理,區塊特徵值包含:區塊內的色彩空間、位置及形狀,本研究目的為減少分割後的影像區塊個數,同時在與影像不失真的權衡下,對全彩影像切割出適切的物件區塊位置。最後利用三種影像分割的績效評估函數進行評比,證實本研究所提方法的兩段式影像分割之方法論,有較佳之結果,優於一般應用SOM於彩色影像分割的單階段處理方法、只考慮色彩特徵的區域合併法,與兩階段啟發式SOM。
Image segmentation technology evolution has a trend toward the three-dimensional color image instead of the one-dimensional gray image. Cluster-based image segmentation algorithms take advantage of color image with more data in three-dimension. However, most cluster-based color image segmentation algorithms discussed in the literature only consider the distribution of a color space, but they seldom consider the features of the image domain. This could result in over-segmentation. Due to this concern this study proposes a two-phase method for the segmentation of color images. In the fist phase, we use the self-organizing feature map (SOM) network method to reduce the information to represent an image by neurons or clustering centers. In the second phase, we merge regions of the segmented image by the region features including color, position, and shape. This research proposes the method to find the suitable number of regions after segmentation; more regions result in over-segmentation, but less regions may distort the original image. This study uses three quantitative evaluation functions of color image segmentation to compare our method with three other methods: the improved single stage SOM method, our method but only color feature is considered, and the two-stage heuristic SOM method, and it is proved that the proposed method has the best performance.